perf(internal): pack from end without cloning input

[fallintoplace]

Summary

• PackEnd now iterates over the input slice backwards instead of cloning and reversing it.
• It reuses the packed outer slice instead of allocating a second [][]T.
• The input slice remains unchanged.
• Empty input keeps returning a non-nil empty result.

Benchmarks

Ran on an Apple M1 Pro with Go 1.26.3. Medians from five runs:

• int64, 128 items: 1.27 us, 3,160 B/op, 12 allocs/op -> 0.96 us, 2,112 B/op, 10 allocs/op
• 64-byte item, 16,384 items: 604 us, 3.16 MB/op, 1,162 allocs/op -> 461 us, 2.11 MB/op, 1,160 allocs/op
• 64-byte item, 131,072 items: 4.41 ms, 25.27 MB/op, 9,229 allocs/op -> 4.04 ms, 16.85 MB/op, 9,227 allocs/op

Tests

• go test -timeout=5m $(go list ./... | rg -v "/io/gocloud$")
• go test -race -timeout=5m ./internal ./table/...
• golangci-lint run --timeout=10m