perf(table): group positional deletes by file path in one pass

[fallintoplace]

What

Group positional deletes by file path in one pass.

The old implementation found the unique paths and then ran an equality expression and filter over the full columns once per path. This version walks the file_path and pos values once, appends each position to its path's builder, and keeps the existing per-file map semantics.

It also:

• checks cancellation before and during large grouping passes
• keeps independent cursors for differently chunked columns
• releases builders if grouping fails
• benchmarks both single-chunk and deliberately mismatched multi-chunk inputs

Benchmark

1M rows, identical generated input on base 65005959 and this branch, Apple M1 Pro, darwin/arm64, -benchtime=100ms -count=3. Values are the median of three runs.

paths	base ns/op	head ns/op	speedup	base B/op	head B/op	base allocs/op	head allocs/op
1	72,511,542	16,228,399	4.5x	60,418,596	17,147,496	2,000,164	39
10	163,576,708	21,973,542	7.4x	62,849,376	21,887,440	2,000,950	317
100	962,813,292	24,432,517	39.4x	87,080,904	29,639,840	2,008,794	2,513
1,000	8,186,696,000	28,130,344	291.1x	327,849,536	19,723,416	2,087,157	17,026

The benchmark also runs 64 file_path chunks against 65 position chunks. In a 500ms, three-run check, this stayed within run-to-run noise of the single-chunk case, with B/op effectively flat and at most one additional allocation per operation.

Checks

• go test ./table -count=1
• go test -race ./table -run '^TestGroupPosDeletesByFilePath' -count=1
• go vet ./table
• golangci-lint run ./table --timeout=10m
• go test $(go list ./... | rg -v '/io/gocloud$') -count=1