Fix and re-enable xfail/skipped tests, clean up stale test TODOs - #15354

curio's module_not_available("curio") check no longer blows up during
collection (it was previously wrapped in an unconditional skip plus a
stale xfail to work around that). Verified with curio 1.6 (latest on
PyPI) that: skip_without("curio") collects fine whether or not curio
is installed, the test passes when curio is installed, and it is
properly skipped (not errored) when curio is absent.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

`test_decorator_skip`, `test_decorator_skip_disabled` and
`test_decorator_skip_with_breakpoint` were all skipped with "recently
fail for unknown reason on CI". The reason turns out to be a Python
version behaviour change rather than flakiness: since Python 3.13,
`set_trace()`/`breakpoint()` stop on the line where they are called
instead of on the following line.

The shared fixture runs

    def f():
        ipdb.set_trace()
        bar(3, 4)

so on 3.13+ the debugger stops on the `ipdb.set_trace()` line and one
extra `step` is needed before the `bar(3, 4)` call is reached. The two
`test_decorator_skip*` tests stepped straight into what they assumed was
the decorated call, landed on `bar(3, 4)` instead, never saw the
`--Call--` frame they expected, and timed out.

`test_decorator_skip_with_breakpoint` already had a version guard, but a
later "try to fix on CI" tweak added a 3.14 branch expecting the current
line to be printed *without* the `-->` marker. That does not match
released 3.14/3.15, where the output is identical to 3.13
(`---> 46     ipdb.set_trace()`); the extra branch is dropped and the
3.13 branch now covers everything from 3.13 on.

Introduce `SET_TRACE_STOPS_ON_CALLING_LINE` to express the behaviour
change once, use it in the three tests, and drop the skip markers.

Verified on CPython 3.11, 3.12, 3.13, 3.14, 3.14t (free-threaded) and
3.15.0b4: all three tests pass, including eight consecutive runs on
3.11/3.13/3.14 and runs under CPU oversubscription.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

`_determine_completion_context` treated any trailing dot as an attribute
access, so nonsense expressions like `3a.` or `$).` were routed to
attribute completion. It also mis-detected nested f-strings (`f'{f'a.`)
because the naive quote tracking sees the inner opening quote as closing
the outer one, leaving `is_string` False while still inside a
replacement field.

Scan backwards from the trailing dot to extract the primary expression it
would apply to (handling bracketed groups and string literals), and only
report an attribute context if that expression parses. Template string
replacement fields are now recursed into whenever we are inside one,
regardless of the `is_string` flag.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>