Prepare shard files ahead of writes by default

[nclack]

Prepare the first shard generation before stream creation returns, then prepare one generation ahead on a separate filesystem worker. Preparation is enabled by default for supported sinks across CPU, GPU, multiarray, and NGFF streams. config.disable_shard_preparation = 1 opts out; unsupported sinks, including S3, retain their existing behavior automatically.

Keep at most one unused prepared file per slot and respect finite stream limits. A private preparation module owns the worker, file adoption, and cleanup, retaining shared-directory ownership across separate cancellations. Cleanup continues across arrays after a failure, preserves existing files and directories, and reports failures. Preparation and ordinary writes share the capacity calculation.

Validation: all 55 non-S3 CPU tests passed, including readback and new lifecycle regressions. Preparation tests also passed with AddressSanitizer and UndefinedBehaviorSanitizer.

[nclack]

This PR is stacked on #288 (borrow-shard-footers) to preserve the tested
footer implementation. Its diff contains the shard-preparation option and
its lifecycle, portability, documentation, and regression coverage.

Preparation uses a separate worker and is excluded from write, metadata,
and footer completion fences. Opening a shard adopts the prepared file;
it waits if preparation has not finished. Flush stops future preparation
before draining delivery, then waits for and removes unused owned files.
Cancellation preserves existing files and directories, reports failures,
and can be retried.

Validation:

• The initial implementation passed 53 CPU tests and all 29 GPU tests.
• After correcting compressed prepared capacity, the final implementation
  passed three focused CPU tests and two GPU tests, including Zarr
  readback and the raw/compressed initial-resize regressions.
• Tests cover empty, partial, complete, finite, and unbounded streams;
  explicit flush/close and destruction; prepared-file adoption; queue
  independence; existing-file preservation; and retryable cleanup.
• Every submitted implementation file matches that final tested version.

[codecov]

Codecov Report

❌ Patch coverage is 68.59504% with 114 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/zarr/shard_preparation.c	68.51%	23 Missing and 28 partials ⚠️
src/zarr/zarr_array.c	59.64%	12 Missing and 11 partials ⚠️
src/zarr/shard_pool_fs.c	71.42%	6 Missing and 8 partials ⚠️
src/multiarray/stream.c	72.22%	2 Missing and 3 partials ⚠️
src/zarr/shard_delivery.c	75.00%	2 Missing and 3 partials ⚠️
src/zarr/io_backend.fs.c	63.63%	2 Missing and 2 partials ⚠️
src/cpu/stream.c	40.00%	1 Missing and 2 partials ⚠️
src/ngff/ngff_multiscale.c	82.35%	1 Missing and 2 partials ⚠️
src/platform/platform_io.posix.c	70.00%	0 Missing and 3 partials ⚠️
src/cpu/stream.body.c	71.42%	1 Missing and 1 partial ⚠️
... and 1 more

❌ Your patch check has failed because the patch coverage (68.59%) is below the target coverage (70.00%). You can increase the patch coverage or adjust the target coverage.

Files with missing lines	Coverage Δ
src/platform/platform_io.h	100.00% <ø> (ø)
src/zarr/shard_write_plan.c	64.39% <100.00%> (-0.03%)	⬇️
src/writer.c	85.00% <80.00%> (-0.46%)	⬇️
src/cpu/stream.body.c	82.35% <71.42%> (-0.41%)	⬇️
src/cpu/stream.c	81.61% <40.00%> (-0.39%)	⬇️
src/ngff/ngff_multiscale.c	64.20% <82.35%> (+1.28%)	⬆️
src/platform/platform_io.posix.c	72.46% <70.00%> (+0.79%)	⬆️
src/zarr/io_backend.fs.c	75.82% <63.63%> (-0.79%)	⬇️
src/multiarray/stream.c	78.22% <72.22%> (+1.59%)	⬆️
src/zarr/shard_delivery.c	73.91% <75.00%> (+0.10%)	⬆️
... and 3 more

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.