perf(table): size partition row indices adaptively

[fallintoplace]

What changed

• Keep the existing maximum starting capacity of 128 row indexes.
• Reduce the starting capacity as more partitions are discovered in a batch.
• Use batch rows / discovered partitions as the estimate, with a minimum of one.
• Keep grouping and row ordering unchanged.

Why

Every new partition currently reserves space for 128 row indexes. This works well for low-cardinality batches, but a batch with one partition per row reserves about 1 KiB for every single row.

The adaptive estimate keeps the current behavior for a few large partitions. It gradually moves toward one row of capacity for high-cardinality batches.

Benchmark

Apple M1 Pro, 32,768 rows, median of 5 runs:

go test ./table -run '^$' -bench '^BenchmarkPartitionRowCapacity$' -benchmem -benchtime=1s -count=5 -cpu=1

Distribution	main	this PR	Change
1 partition	1,160,120 B/op	1,160,120 B/op	unchanged
16 partitions	951,784 B/op	951,784 B/op	unchanged
1,024 partitions	1,759,064 B/op	1,348,728 B/op	-23%
1 partition per row	50,331,068 B/op	18,241,259 B/op	-64%
90% in one partition	5,994,673 B/op	3,580,624 B/op	-40%

Allocation counts are unchanged in every case. The unique-partition case was also about 14% faster in the repeated comparison.

Tests

• go test ./...
• go test -race ./table -run '^TestFanoutWriter$' -count=1
• go vet ./...
• git diff --check