Optimize startup performance by deferring non-critical imports - #15353

`refresh_style()` turns a whole pygments style into prompt_toolkit style
rules, and is called several times while a shell is set up: once from
`init_syntax_highlighting` when `colors` is first set, again from
`init_magics`, again when the prompt_toolkit application is created. Only the
last of those states is ever observed, and for a run that never draws a
prompt -- `ipython -c ...`, `--simple-prompt`, a kernel, the test suite --
none of them are.

Make `refresh_style()` invalidate a cache instead, and build the style behind
a `_style` property. `_style` is only ever read through the `DynamicStyle`
the prompt session renders with, so the rebuild lands on the first render.
The property keeps a setter so that assigning to `_style` still works.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

`IPython.terminal.shortcuts.filters` evaluates the filter expressions of the
built-in key bindings at import time, and needed three names from
`IPython.core.guarded_eval` to do it: `BINARY_OP_DUNDERS`, `UNARY_OP_DUNDERS`
and `_find_dunder`. Importing `guarded_eval` for them cost ~5ms and pulled in
`typing_extensions` on every terminal start, for two dicts and a six-line
helper that only need `ast`.

Move those (and `COMP_OP_DUNDERS`, which belongs with them) to
`IPython.core._dunder_ops`, and re-export them from `guarded_eval`, which
stays their documented home.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

Four more stdlib imports that every `import IPython` paid for and almost no
session used:

- `platform` in `IPython.utils.sysinfo`, needed only by `pkg_info()`;
- `pprint` in the same module, needed only by `sys_info()`;
- `textwrap` in `IPython.core.interactiveshell`, needed only to dedent the
  banner used when `SOURCE_DATE_EPOCH` is set;
- `html` (and with it `html.entities`) in `IPython.core.oinspect`, needed
  only to build the html mime representation of an inspected object;
- `platform` in `IPython.lib.pretty`, needed only to special-case PyPy when
  pretty-printing a `super` object.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

The kitty graphics autodetection walks the process tree looking for a
terminal that speaks the protocol. It runs on every interactive start -- a
headless `ipython -c ...` returns early on the isatty check, which is why
this never showed up in startup profiles -- and importing psutil for it costs
~13ms and 6 modules.

On Linux, ``/proc/<pid>/stat`` holds both the name psutil would report and
the parent pid, so one read per ancestor is all the walk needs. Everywhere
else -- macOS, or a Linux without ``/proc`` -- fall back to psutil as before.
`comm` in that file is unescaped and may contain spaces and parentheses, so
the name is taken from the last ``)``; it is also truncated to 15 characters,
which cannot affect the outcome since every terminal name matched against is
far shorter.

Also replace `platform.system()` with `sys.platform`, which answers the same
question without importing `platform` (another ~1.5ms).

Together this takes 48 modules off an interactive start.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

Four more single-use stdlib imports on the `import IPython` path:

- `mimetypes` (and the `winreg`/`_winapi` probing it does) in
  `IPython.core.display`, used only to guess the type of an embedded video;
- `locale` in `IPython.utils.encoding`, consulted only when the stream
  encoding is missing or ascii -- which it usually is not, so this was an
  import for a call that mostly never happened;
- `glob` in `IPython.utils.path`, used only by `shellglob()`;
- `runpy` in `IPython.core.interactiveshell`, used only by
  `safe_run_module()`.

Making `mimetypes` lazy surfaced that `tests/test_path.py` registers a stub
`winreg` module in `sys.modules` and never removes it, so any stdlib module
importing `winreg` later in the session -- `mimetypes` does, to decide
whether to read the Windows registry -- found the stub and tried to use it.
The only test that needs `winreg` runs on Windows only, so import it there
and leave `sys.modules` alone everywhere else.

The encoding tests reached `locale` through `IPython.utils.encoding.locale`;
patch the `locale` module itself, which is the same object either way.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR

`test_supports_kitty_graphics_does_not_import_psutil_on_linux` set
`sys.platform = "linux"` and asserted the walk answers without psutil, but on
the Windows and macOS runners there is no `/proc`, so detection correctly fell
back to psutil and the test failed -- the only failure in the six red jobs.

Skip it unless `/proc/self` is really there, and stop faking `sys.platform`:
what the test is for is that a real Linux answers without psutil. The psutil
branch is still covered on every platform by
`test_supports_kitty_graphics_handles_psutil_access_denied`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9XqGGqj9CZhSXWJ49t7fR