perf(distributed): skip eager probe session when chunkWorkerCount > 1

[jrusso1020]

What

When renderChunk resolves chunkWorkerCount > 1, skip the eager
createCaptureSession + assertSwiftShader + initializeSession pre-warmup
that captureStage's parallel branch immediately closes anyway. Move the
SwiftShader safety probe into executeWorkerTask so each parallel worker
validates its own GPU backend before its first frame.

Why

The OSS comment at the probeSession: session callsite already flagged
this:

// The parallel branch closes this session and spins up its own
// worker sessions, wasting the ~3-5s of pre-warmed setup. Worth a
// follow-up to skip pre-warmup when the resolved workerCount > 1.
probeSession: session,

In the dev distributed-render benchmark (12 producer-worker pods, v0.6.15
sidecar), texture-launch's chunk_p95 lands at ~25-35s with most of the
non-capture time being per-chunk fixed overhead. ~3-5s of that is the
probe pipeline we throw away — meaningful at the small chunks that real
maxParallelChunks settings produce.

How

• packages/engine/src/utils/readWebGlVendorInfoFromCanvas.ts — new file.
  Moved from producer/services/distributed/renderChunk.ts; both
  producer and engine now need it. renderChunk.ts keeps a re-export
  via export { readWebGlVendorInfoFromCanvas } from "@hyperframes/engine";
  so @hyperframes/producer/distributed's public surface is unchanged
  (the existing publicExports.test.ts assertion still passes).

• packages/engine/src/services/parallelCoordinator.ts —
  executeWorkerTask now runs
  assertSwiftShader(session.page, readWebGlVendorInfoFromCanvas)
  after createCaptureSession when config.browserGpuMode === "software".
  Each parallel worker self-validates SwiftShader before its first
  frame.

  In-process renders default to browserGpuMode: "software", so they
  also pick up this safety net. Cost is one about:blank navigation +
  one page.evaluate per worker (~100-200ms, concurrent across workers
  so wall-clock impact ≈ slowest worker probe). Hardware-GL paths
  (browserGpuMode !== "software") are untouched.

• packages/producer/src/services/distributed/renderChunk.ts —
  resolve chunkWorkerCount up-front, skip the entire pre-warmup
  branch when chunkWorkerCount > 1, pass probeSession: null in
  that case. Sequential path (chunkWorkerCount === 1) is unchanged:
  it still pre-warms because captureStage's sequential branch reuses
  the probe.

Test plan

• Unit tests added/updated — parallelCoordinator.test.ts's
  existing distributeFrames/calculateOptimalWorkers coverage is
  unaffected; behavior gating is small enough to verify by reading.
• Existing tests pass:
  • packages/engine — all 605 tests pass (bun run test).
  • packages/producer distributed unit tests — 47 pass
    (assemble, plan, publicExports, chunkBoundary,
    planFormatBanlist, planSizeCap).
• Lint + format — bunx oxlint and bunx oxfmt --check clean
  across the four touched files.
• Docker-based regression renderChunk.test.ts byte-identical-retry
  test — runs the fixture at chunkWorkerCount=1 (5 frames), exercising
  the sequential path which is unchanged. The parallel path will be
  exercised by re-benchmarking on dev after release.
• Re-benchmark texture-launch on dev after v0.6.16 release. Expect
  chunk_p95 to drop ~3-5s at chunks ∈ {3, 6, 8} (where
  chunkWorkerCount > 1 and the probe is currently wasted). chunks=12
  hits chunkWorkerCount=1 and is unchanged.

Notes for reviewers

• The flagged // follow-up comment in
  services/distributed/renderChunk.ts is removed.
• readWebGlVendorInfoFromCanvas is unchanged in behavior — just
  relocated.
• I considered a per-task assertSoftwareGl?: boolean flag instead of
  gating on cfg.browserGpuMode === "software", but the latter matches
  the existing safety contract semantics cleanly: "if you declared
  software GL, we verify it." Open to flipping if preferred.