feat(rescheduling): optional crossPool mode for gpuFragmentation

[pfernandes21]

What type of PR is this?

/kind feature

What this PR does / why we need it:

Adds an opt-in crossPool flag (default false) to the gpuFragmentation rescheduling strategy. Today candidate sources/destinations are grouped per karpenter.sh/nodepool, so consolidation never crosses a pool boundary and unlabeled (e.g. on-prem) nodes are invisible.

With crossPool: true:

• every GPU node in the cluster forms one candidate set, including nodes without the pool label;
• the workload's own node selection — nodeSelector, required affinity, taints/tolerations — evaluated through the existing session predicate path, is the only placement filter;
• the cooldown clock, one-drain-per-pass budget, and metrics pool label become cluster-wide (label value cross-pool).

Planner change is confined to the grouping loop:

if conf.CrossPool {
    pools[crossPoolName] = append(pools[crossPoolName], node)
    continue
}
pool, ok := node.Node.Labels[conf.PoolLabel]
...

Everything downstream (FFD simulation over cloned idle capacity, fractional-fullness destination ordering, atomic move sets, priority ceiling/tie-break, per-PodGroup eviction caps, stamping) is unchanged. Default behavior with crossPool: false is byte-for-byte the existing per-pool behavior.

Which issue(s) this PR fixes:

Fixes #

Special notes for your reviewer:

• New tests: cross-pool drain that per-pool mode blocks, unlabeled-node destination, and a predicate veto blocking an otherwise-fitting cross-pool move; all existing per-pool tests unchanged and passing.
• Predicates are the only guard on placement: pods with loose/no node selection may move across instance types, capacity types, or AZs. That is the explicitly chosen semantic.
• go test ./pkg/scheduler/plugins/rescheduling/... passes; go vet/gofmt clean.

Does this PR introduce a user-facing change?

gpuFragmentation gains an opt-in crossPool option: when true, GPU repacking considers all GPU nodes cluster-wide (including nodes without the pool label) and relies solely on workload node selection via scheduler predicates for placement; cooldown and drain budget apply cluster-wide. Default remains per-pool.

Link to Devin session: https://app.devin.ai/sessions/9f5f310b988c4566869471b54b1154dd
Open in Devin Desktop: https://app.devin.ai/desktop/session/9f5f310b988c4566869471b54b1154dd?variant=devin
Requested by: @pfernandes21

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring