Persist the numba compile of the built-in UDF binops to disk - #611

floordiv, rfloordiv, absfirst, abssecond, and rpow are UDFs, so every
process that touched one paid its numba compile again. They are plain
module-level functions, which is exactly the case numba can cache, so
that cost belongs once per machine rather than once per process.

Thread the flag through as a private keyword-only _cache on
register_new into _build, where it reaches numba.njit alongside the
existing error_model. Keeping it private avoids widening the public API
for something no caller outside this module should set. Ops registered by
users stay uncached: numba keys a cache entry on a stable on-disk source,
which lambdas and interactively defined functions do not have.

Measured here, fresh process, medians of 3, load average ~4 on a shared
machine so read these as indicative. First-touch floordiv alone: 434 ms
uncached, 347 ms warm (-20%). First-touch of all five: 1337 ms uncached,
733 ms warm (-45%). Writing a cold cache costs about 107 ms once for all
five (1444 ms on the run that populates it).

What stays: the remaining ~733 ms is numba/LLVM startup and per-process
object linking, which no on-disk cache removes. In-process dispatch once
an op is built is untouched. The first process on a machine is slightly
slower, not faster.

Read-only installs are safe, verified rather than assumed. With
graphblas/core/operator and its pycache chmod'd a-w, the run wrote
zero files in-tree and 66 .nbi/.nbc files under ~/Library/Caches/numba,
with empty stderr under python -W always. The NumbaWarning in
numba/core/caching.py is a source-content check, not a filesystem one.

docs/user_guide/udf.rst gains a "Compilation caching" section covering the
cache location, the read-only fallback, and why user ops are excluded.

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}