feat(driver): expose per-shard io-wq registration results

[houseme]

Related Issues

Related to rustfs/backlog#2647, step 4.2.iowq. This reports the existing io-wq startup request; it does not change worker limits or close the application-wide budget item.

Summary of Changes

• Capture each shard's existing [16, 0] io-wq registration request and result during startup.
• Expose shard_iowq_setup() in shard order. Successful registration records the kernel-returned previous limits; failure records the I/O error kind and optional errno.
• Keep best-effort startup, probing, reads, and fallback behavior unchanged. The query copies stored values without a new kernel call or per-read work.
• Add injected success/error tests, a native two-shard read and snapshot test, and an operational contract.

Verification

• Linux-target cargo check --locked --all-targets --all-features passed.
• Linux-target cargo clippy --locked --all-targets --all-features -- -D warnings passed.
• Default library check and default/all-feature warning-denying rustdoc passed.
• cargo fmt --all --check and git diff --check passed.
• Two independent reviews of the complete code, tests, and documentation found no blocking issue.
• CI at 193ab21 passed all three jobs: 147 native all-feature tests, mandatory O_DIRECT, ordered-prefetch and both benchmark correctness smokes, and restricted/unrestricted Docker coverage. The new three injected unit tests and two-shard native test executed with no skip. The documentation-only final head 2b79358 also passed all three CI jobs.

Impact

Registration still requests sixteen bounded workers per NUMA node and leaves the unbounded class unchanged. Failure remains nonfatal. The recorded previous limits are historical values, not effective limits or active worker counts; no process-wide cap or performance gain is claimed. Existing constructors and dependency features are unchanged.

Additional Notes

This PR is independently revertible. No data or configuration migration is required. Application dependency wiring, result memory limits, and resource-isolated performance acceptance remain separate work.