perf(telemetry): stop copying JSON payloads when converting to Parquet rows

[vishr]

Fifth item of #254.

Problem

Span, Log and Metric held their JSON columns as []byte, and makeSpanParquetRow and its siblings converted each with string(...):

ResourceJSON: string(r.ResourceJSON), AttributesJSON: string(r.AttributesJSON),
EventsJSON:   string(r.EventsJSON),   LinksJSON:      string(r.LinksJSON),

The originals stay live until the commit is durably acknowledged, so every batch held its JSON twice — and four commit workers run at once, so it was four duplicate copies resident at exactly the moment the process is most likely to be killed.

Change

The fields and the ingest producers that fill them are strings, so conversion assigns rather than copies.

attrsJSON already copied out of its pooled buffer before returning it (append([]byte(nil), buf.Bytes()...)), and buf.String() is the same single allocation — so the producers are no worse. The second copy downstream is what goes away.

Measurement

20,000 spans carrying 8.1 MiB of JSON, measuring the conversion alone:

	allocated
before	18.2 MiB
after	9.6 MiB

9.6 MiB is the row-struct headers and nothing else. The JSON is no longer duplicated.

Being straight about the end-to-end number

The commit-level gate does not move outside its own noise. It spans 2.8x-5.6x payload run to run, because parquet-go and zstd pool their buffers, and a saving this size hides inside that spread. Before: 3.24x-5.31x. After: 2.79x-5.64x. Those are the same distribution.

That is why this PR adds TestSpanConversionDoesNotCopyJSONPayloads, which measures the conversion in isolation, where the signal is clean. commitAllocRatioLimit stays at 7.0 - I am not ratcheting it on a change I cannot show through it.

The saving is real and it compounds in production in a way the synthetic test cannot show: the duplicate was held per in-flight commit, across four workers, with more batches queued behind them.

Scope

[]byte to string across internal/telemetry/rows.go, the seven producers in internal/ingest/server.go, parquet_rows.go, their tests, and the experiments/ fixtures that share the module.

No behaviour change. makeSpanParquetRow already turned a nil slice into an empty string, so an absent column is written exactly as before.

just check passes, go vet ./... clean.