feat(ublk): optional Discarder and ZeroWriter capabilities

[ValentaTomas]

DISCARD and WRITE_ZEROES are distinct in the block layer (discard leaves the range undefined; write-zeroes must read back as zeros), so they're exposed as separate optional Backend capabilities — implement either or both:

type Discarder interface {
    DiscardAt(off, length int64) (int, error)
}
type ZeroWriter interface {
    WriteZeroesAt(off, length int64, flags ZeroFlags) (int, error)
}

ZeroFlags carries ZeroNoUnmap (UBLK_IO_F_NOUNMAP, kernel bit 15) so backends that want to keep the range physically allocated can do so. UBLK_PARAM_TYPE_DISCARD (max_discard_sectors / max_write_zeroes_sectors) is advertised per capability, so the kernel only issues an op the backend can serve. The DISCARD / WRITE_ZEROES path bypasses the data-plane buffer and isn't capped by WithMaxIOSize since neither op transfers user data.

[cursor]

PR Summary

Medium Risk
Touches the IO hot path and kernel parameter advertisement; mistakes can surface as kernel I/O errors or unexpected op dispatch under real filesystems. Coverage is improved with new unit and integration tests, but behavior depends on kernel expectations for DISCARD/WRITE_ZEROES semantics.

Overview
This change adds optional Discarder and ZeroWriter backend capabilities and wires them into worker.handleIO for OP_DISCARD and OP_WRITE_ZEROES, including ZeroNoUnmap flag plumbing.

Device.setParams now conditionally advertises discard/zeroing limits only when the backend implements the matching interface, but OP_FLUSH remains unhandled and large discard/zero requests can now bypass the data-plane buffer size constraints.

^{Reviewed by Cursor Bugbot for commit ebd5deb. Bugbot is set up for automated code reviews on this repo. Configure here.}

[ValentaTomas]

Refactored per feedback:

• Split ZeroWriter into two interfaces — Discarder (DiscardAt) and ZeroWriter (WriteZeroesAt(off, length, flags)). Same semantic split the block layer uses (DISCARD = data undefined; WRITE_ZEROES = reads must return zeros), so a backend can implement them independently or together.
• Added ZeroFlags carrying ZeroNoUnmap (UBLK_IO_F_NOUNMAP, kernel bit 15). The orchestrator can ignore it today; exposing it here means we don't silently drop the kernel signal.
• Per-capability params: max_discard_sectors / max_write_zeroes_sectors set only when the corresponding interface is implemented.
• Compacted tests: 4 unit + 4 integration → 2 unit + 2 integration. Dropped the BLKZEROOUT-without-ZeroWriter case since the kernel falls back to plain WRITE there (not observable from an ioctl).

[ValentaTomas]

Both points in the Bugbot summary are intentional:

• OP_FLUSH unhandled — out of scope for this PR. We don't advertise UBLK_ATTR_VOLATILE_CACHE, so the kernel won't emit OP_FLUSH under normal use. Tracked as the new "Handle OP_FLUSH explicitly" entry in TODO.md.
• DISCARD / WRITE_ZEROES bypass the data-plane buffer — by design. Neither op transfers user data (the kernel passes only offset + length), so capping at WithMaxIOSize (the read/write buffer) would artificially limit pure-metadata ops. The advertised cap is 1 << 21 sectors = 1 GiB per op, and TestBlkDiscardAndZeroOut exercises a 1 MiB discard explicitly to prove the no-buffer path isn't capped by bufSize.