go: cut Rows' remaining per-call allocations (#456)

[EricAndrechek]

refs #456

#462 made the decoder allocation-light; this cuts what a full Schema.Rows call still allocated around it. Go only; scope is go/chtypes/**, go/internal/abi1/take.go and go/CHANGELOG.md. No exported name changes, no generated file edited, stock encoding/json only.

Measured (darwin-arm64, Go 1.27.1, the reporter's call shape: 100 NDJSON rows, tenant = {p:String} filter with params, WithExport(JSONCompactEachRow), WithDocFlags(0), a fresh four-key settings map each call)

per call	before (#462)	after
Schema.Rows	412 allocs, 75.2 KB	76 allocs, 45.4 KB
same, settings map built once by the caller	409, 74.9 KB	74, 45.1 KB

Allocation sites, from -memprofile -memprofilerate 1, alloc_space / alloc_objects, per call:

site	before	after	notes
rows result slice ([]RowResult)	23.4 KB / 1	23.4 KB / 1	returned to the caller; only an API change shrinks it
document copy out of the library	16.6 KB / 1	0 / 0	pooled buffer (TakeInto)
export payload copy	3.5 KB / 1	3.5 KB / 1	returned to the caller, stays a fresh Go slice
json.Decoder buffer growth (stock)	11.9 KB / 13	8.1 KB / 7	the stock decoder's growth policy; the 2.4 KB row_spans list is no longer read as one value
per-row Verdict pointers	1.6 KB / 100	2 KB / 4	shared chunks of 32
PartitionID pointer (allocated even when the row has none)	1.6 KB / 100	0 / 0	one chunk when a row has one
per-row InputSpan pointers	2 KB / 4	2 KB / 4	already chunked
row_spans decode ([]*spanDoc, reflect, growslice, spansOf)	7.2 KB / 117	1.9 KB / 3	streamed element by element into the []Span the result carries
decoder locals (top-level fields)	0.4 KB / 17	0.4 KB / 2	one struct
row decoder, sub-decoder	0.5 KB / 3	0 / 0	pooled
settings JSON	0.5 KB / 10	0 / 1	pooled encoder, result cloned
option closures	5 allocs	0	small named types
callConfig, generated Buf handles	1 + 4	1 + 4	generated code and the call itself; not cuttable here
caller's own settings map	0.28 KB / 2	same	not the SDK's

Safety

• Nothing returned aliases C memory or a pooled buffer: TakeInto copies, and TestDecodeBatchOwnsItsMemory overwrites the document buffer after decoding and compares.
• Rows output is identical to go: allocation-light Rows decoding (#456) #462's across doc_flags 0, 1, 2, 3, 7, 15, 31, 255, body sizes 0, 1, 7, 100, 300 and a malformed body (30 cases, JSON-hashed, run against both trees with the real library; local, not in CI).
• TestDecodeBatchEquivalence* stay and gain a third real document (a filtered call, with row_spans, rows_passed and rows_cut) and 14 row_spans mutations.
• go test -race ./chtypes passes (stub-driven concurrent Rows test in CI; a real-library 8-goroutine Rows check under -race run locally).

Not in this PR: anything that needs an ABI or library change (the generated Buf wrappers and their finalizers, the []RowResult shape, the stock decoder's own buffer growth).

🤖 Generated with Claude Code