perf: use stdlib bisect and attrgetter in tablets.py

[mykaul]

Summary

• Use stdlib bisect.bisect_left on Python >= 3.10 (C implementation) instead of the bundled pure-Python copy. The fallback is retained for Python < 3.10 where key= is not supported.
• Replace per-call lambda closures with module-level operator.attrgetter for first_token / last_token extraction, avoiding repeated function-object allocation.
• Fix bug in the pure-Python bisect_left fallback: the key=None branch executed bare return (returning None) instead of return lo. This bug is currently latent (all callers pass key=), but would cause silent incorrect behavior if bisect_left were ever called without key.
• Add unit tests for the fallback bisect_left (7 tests) and get_tablet_for_key (3 tests).

Benchmark Results

Measured on Intel i7-1270P, Python 3.14.3, pytest-benchmark.

get_tablet_for_key (hit — the primary hot path)

Tablets	Before (ns)	After (ns)	Speedup
10	517	365	1.42x
100	616	351	1.75x
1,000	1,008	529	1.91x
10,000	1,339	610	2.20x

bisect_left with key= (isolated)

Size	Before (ns)	After (ns)	Speedup
10	278	216	1.29x
100	450	289	1.56x
1,000	697	406	1.72x
10,000	1,043	496	2.10x

bisect_left without key= (plain ints — bug fix path)

Size	Before (ns)	After (ns)	Speedup
10	162	56	2.89x
100	268	73	3.67x
1,000	495	96	5.16x
10,000	717	117	6.13x

The full TokenAwarePolicy.make_query_plan with tablet lookup shows minimal change because it is dominated by other costs (policy iteration, host filtering), but the get_tablet_for_key component is substantially faster.