fix(scheduler): skip devices with malformed MIG UUID in getNodesUsage

[Wangmin362]

What type of PR is this?

/kind bug

What this PR does / why we need it:

getNodesUsage discards the error returned by ExtractMigTemplatesFromUUID and uses the returned indices without checking them. A pod carrying a malformed or corrupt MIG UUID annotation therefore panics with index out of range:

• GPU-MIG33[999] (no - inside the brackets): ExtractMigTemplatesFromUUID returns an error and -1 indices; the discarded error lets -1 reach PlatternMIG (templates[-1]) and MigUsage.UsageList[Instance].
• GPU-MIG33[5-0] (well-formed but the template index does not exist on the node): ExtractMigTemplatesFromUUID returns (5, 0, nil) — no error — and PlatternMIG then indexes templates[5] out of range.

getNodesUsage runs in the background register loop started by main.start — a goroutine with no HTTP recover — so any of these panics kills the whole scheduler process and puts it into CrashLoopBackOff. Any pod with such an annotation can repeatedly crash the scheduler.

This PR handles it in three places:

• skip the device when ExtractMigTemplatesFromUUID returns an error (mirrors the existing handling in pkg/device-plugin/nvidiadevice/nvinternal/plugin/util.go);
• bound-check the instance index before indexing MigUsage.UsageList;
• bound-check the template index inside PlatternMIG so an out-of-range index returns instead of indexing templates[idx] (this also protects the other PlatternMIG callers).

Which issue(s) this PR fixes:
NONE

Special notes for your reviewer:

Real-machine before/after verification in an isolated single-node kind cluster (its own apiserver, zero interaction with any shared/production cluster), using scheduler images built from this branch:

• Before (unfixed): applying a pod with hami.io/vgpu-devices-allocated: "GPU-MIG33[999],NVIDIA,4096,0:;" assigned to a MIG node crashes the scheduler within ~8s (RESTARTS 0->1, CrashLoopBackOff), with the panic stack index out of range [-1] at PlatternMIG <- getNodesUsage <- register <- RegisterFromNodeAnnotations, created by main.start.
• After (fixed): both the GPU-MIG33[999] (malformed) and GPU-MIG33[5-0] (out-of-range template index) pods are processed every register tick and skipped with a logged error; the scheduler stays up (RESTARTS 0) across multiple ticks.

Unit regression test added (Test_getNodesUsage_MalformedMIGUUID, three sub-cases: malformed UUID, out-of-range instance index, out-of-range template index): each panics before the corresponding guard and passes after; full pkg/scheduler and pkg/device suites pass.

AI assistance: the analysis is based on my own testing; the code was drafted with AI (Claude Code) assistance and reviewed and verified by me.

Does this PR introduce a user-facing change?:

Yes - the scheduler no longer crashes when a pod carries a malformed or out-of-range MIG UUID annotation; the offending device is skipped instead.

[Wangmin362]

Fixed upstream by Project-HAMi#2088 (bound-check MIG template/instance index parsed from UUID annotation). Closing.