Context

While working on PYTHON-5965, I found a performance win that affects the existing
bson. I measured about a 15% improvement on a spawn perf host on the standard bson tests. I measured a 52% encode and 19% decode performance increase for datetime objects using the  Howard Hinnant algorithm, which is MIT licensed.

Definition of done

• Replace the time64 library's loop-based calendar math with Hinnant's O(1) algorithm: cbson_timegm64 (encode) iterates year-by-year from 1970 (~8000 iterations for year 9999), and cbson_gmtime64_r (decode) loops per-year.
• Swap millis_from_datetime/datetime_from_millis over to days_from_civil/civil_from_days and drop the time64 sources.
• Remove bson/time64.c|time64.h|time64_config.h|time64_limits.h
• Verify no regression in round trip values

Pitfalls

We need to verify calendar-math correctness across year 1..9999 and negative/epoch boundaries.
Compare the new implementation against the known-good reference across many inputs, AND verify data survives a full encode → decode loop.
Also verify that test_bson_corpus.py passes.