feat(query-backend): hedge GetRange calls to reduce symbol fetch tail latency

[simonswine]

Summary

Symbol table fetches (rawTable[T].fetch, stacktraceBlock.fetch) issue GetRange calls against object storage and are the primary tail latency hotspot on the query backend read path. A single slow fetch stalls the entire fetchTx for a block.

This PR adds opt-in speculative hedging: after --query-backend.block-read-hedge-after, a second parallel GetRange is issued against the same object. Whichever responds first is decoded; the other is cancelled and its connection closed.

Design decisions:

• Hedge only GetRange, not the decode — the winning RC is decoded once (no duplicate CPU work)
• New Cleanup func(T) field on retry.Hedged closes the losing io.ReadCloser to prevent connection leaks when both calls succeed (see test(retry): use synctest for deterministic hedging tests #4975 for the synctest test cleanup that's a prerequisite)
• Disabled by default (hedgeAfter=0); opt-in per deployment

Changes

• pkg/util/retry/hedged.go — add Cleanup func(T) field; called on a losing attempt's result when it succeeded but the other won
• pkg/phlaredb/symdb/block_reader.go — add hedgeAfter to Reader; extract getRange() helpers on rawTable[T] and stacktraceBlock that wrap GetRange with retry.Hedged when configured
• pkg/block/object.go — add hedgeAfter field + WithObjectHedgeAfter option
• pkg/block/section_symbols.go — thread hedgeAfter into symdb.OpenObject
• pkg/querybackend/block_reader.go — add WithBlockReaderHedgeAfter option; pass to block.NewObject
• pkg/querybackend/backend.go — add BlockReadHedgeAfter config field + flag

Depends on

• test(retry): use synctest for deterministic hedging tests #4975 — synctest cleanup (the Cleanup field tests use synctest)

Test plan

• go test ./pkg/util/retry/... ./pkg/phlaredb/symdb/... ./pkg/querybackend/...
• Enable with --query-backend.block-read-hedge-after=2s and verify GetRange is issued twice for slow fetches

---

Note

Medium Risk
Introduces new speculative parallel GetRange calls and changes retry.Hedged cancellation/cleanup semantics, which can affect concurrency behavior and resource usage under load. Disabled by default but could increase object-store traffic when enabled and needs careful tuning.

Overview
Adds opt-in hedged object-store reads to reduce tail latency when fetching symbol tables. A new query-backend.block-read-hedge-after setting is threaded from querybackend.Config into BlockReader/block.Object and down to symdb.OpenObject, enabling a second speculative GetRange after the configured delay.

pkg/phlaredb/symdb now wraps symbol table and stacktrace GetRange reads with retry.Hedged (including closing the losing io.ReadCloser). retry.Hedged itself is updated to use per-attempt contexts (so the winning result remains usable) and adds a Cleanup hook, with new tests covering winner-context behavior and cleanup semantics.

^{Reviewed by Cursor Bugbot for commit 1a87ad3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before autofix could start.}

[aleks-p]

LGTM