fix(linstor): wait for temporary probe devices - #4100

LGTM (independent cozy-review pass).

Vendored Java patch to piraeus-server, origin = upstream LINBIT PR #528 / fixes #527. It applies cleanly alongside the other five patches (including the sibling luks patch that touches adjacent lines in the same file) and the result is structurally sound. The wait loop is bounded — attempts < PROBE_VLM_ATTEMPTS (3), waitUntilDeviceCreated uses a bounded device-visible timeout that throws on expiry, the 1s retry delay is interruptible, cleanup runs in a finally each iteration, worst case ~17s — no unbounded busy-wait. The LVM path is explicitly preserved (probe only on empty list or ZFS/ZFS_THIN), and a new null-path guard fixes a latent NPE.

Non-blocking: the bundled JUnit test never compiles/runs in the cozystack image build (gradle installdist has no test task), so there is no automatic regression signal here — recommend a live validation on a 4K ZFS pool.

LGTM. Checked the committed diff against LINBIT/linstor-server#528 hunk by hunk and applied the whole patches directory to a clean v1.33.3 checkout. All six apply in the order the build globs them, this one lands at its exact offsets, no fuzz.

Business context: LINSTOR reads a ZFS probe volume's block device properties before udev publishes the device node, so 4K pools end up with no min/opt I/O size properties and replica placement fails.

Most of what differs from #528 is forced by 1.33.3. The dropped CHANGELOG.md, javadoc and LvmProviderBlockDeviceInfoTest.java hunks have nothing to attach to at this tag. The null-path guard is missing from #528 because upstream already had it, and it is needed here or haveInfo goes true on a null path and the ZFS fallback never runs. The ZFS restriction has a better reason than the README gives: upstream can probe every provider because it tries getReadOnlyProbeDevice() first, and 1.33.3 has no such method, so an unrestricted fallback would create a volume on an inactive shared LVM volume owned by another node. Two deviations that are not forced are in the list below.

No JDK 11 or Gradle on my side, so the 1026 tests, assemble and "14 of 17 fail without the patch" stay unverified. The 17 regression tests are in the diff. DrbdAdm.java:970 is the for loop in secondaryAfterMkfs, which retry-secondary-after-mkfs.diff adds and which returns from inside the loop, so that attribution holds. Nothing in CI builds this image, the four green checks are labels, size, DCO and the bot.

Four things for later, none of them blocking:

1. The new test reads BlockSizeConsts.DFLT_OPT_IO_SIZE, which fix-luks-header-size.diff adds. It compiles today only because the glob sorts luks first. Drop luks after some future bump and this test breaks with a missing symbol nowhere near the file that went away, so the coupling is worth a line in the README entry.
2. Upstream gates cleanup on a created flag and gives up after a failed creation. Here delete always runs and the loop retries. With the real provider it never gets that far: ZfsCommands.delete shells out to zfs destroy with only a device-busy retry handler, so destroying a volume that was never created throws, cleaned stays false and the loop breaks on the first pass. Costs an extra zfs destroy and an extra warning. failedCreationStillCleansEveryAttempt and transientCreationFailureIsRetried pin behaviour only the mock produces.
3. interrupted looks at exc.getCause() only. Upstream also checks Thread.currentThread().isInterrupted() and has a test for it that this patch drops. The device wait wraps InterruptedException as the cause so the common path is covered, but an interrupt during the block size read costs an extra iteration and the final warning then blames the interrupted delay instead of the real failure.
4. No removal note. #528 merged as 6e55578 and v1.35.1 contains that commit, so "drop when the image moves to LINSTOR 1.35.1 or newer" is nameable. It is also the only entry with no backported commit, and it sits after retry-secondary-after-mkfs instead of the alphabetical order the rest of the file and the glob both follow.