Make partition expression evaluation stateless

[dossett]

Partition pruning currently rebuilds and binds the same expression for every data file because _ExpressionEvaluator stores the current record on itself and cannot safely be shared across manifest worker threads. Even for a single equality predicate, repeated binding costs more than the actual partition evaluation work.

This change separates immutable expression preparation from call-local evaluation state. That makes one prepared evaluator safe to share for every manifest using a partition spec, without locks or thread-local state. It is a fresh alternative to apache#3656, rooted directly at Apache main rather than stacked on that branch.

Summary

• Keep the bound expression on _ExpressionEvaluator and move each record into a short-lived visitor.
• Prepare one partition evaluator per spec and share it across manifest tasks.
• Add deterministic concurrency, immutable-state, and planner-sharing coverage.
• Add a benchmark using a single equality predicate across dense and fragmented manifests.

Performance

The focused benchmark evaluates 1,000 files against x == 5 using a two-field partition spec. Each result is the mean of five runs, with each run averaging 100 iterations.

Revision	1,000 files/manifest	1 file/manifest
main	5.001ms	5.199ms
Task-local reuse (apache#3656)	0.671ms	5.312ms
This change	0.756ms	0.928ms

Compared with main, stateless per-spec sharing is about 6.6x faster for a dense manifest and 5.6x faster for fragmented manifests. Compared with task-local reuse, constructing a call-local visitor is about 13% slower in the dense case, but avoiding one bind per manifest is about 5.7x faster in the fragmented case.

This is a focused evaluator benchmark, not an end-to-end query-runtime claim. It excludes catalog access, manifest Avro I/O and decompression, metrics evaluation, residual planning, and downstream data reads.

Testing

• .venv/bin/python -m pytest tests/expressions/test_visitors.py -q
• .venv/bin/python -m pytest tests/expressions/test_evaluator.py -q
• .venv/bin/python -m pytest tests/table/test_init.py -q
• .venv/bin/python -m pytest tests/table/test_partition_evaluator_planning.py -q
• .venv/bin/python -m pytest tests/benchmark/test_partition_evaluator_benchmark.py -v -s -m benchmark
• UV_NO_CONFIG=1 .venv/bin/prek run -a