[Docs] Document stage worker sizing and with_ backend overrides #2129

Context

Child of #2118. #2077 moved Ray Data actor stages to ActorPoolStrategy and added min/max/initial worker sizing. #2086 made ray_stage_spec, xenna_stage_spec, and num_workers configurable through ProcessingStage.with_(). #2110 changed CPU video stage reservations to enable Ray Data fusion.

Required coverage

• Fixed actor pools via num_workers()
• Autoscaling actor pools via MIN_WORKERS, MAX_WORKERS, and INITIAL_WORKERS
• Ray Data task-stage TaskPoolStrategy
• Precedence/warnings when fixed and autoscaling settings conflict
• Xenna cluster-wide worker counts vs num_workers_per_node and invalid combinations
• Shallow-merge and explicit-None semantics for with_() stage-spec overrides
• Reserved num_workers hook and ASR dataloader_num_workers rename
• Resource matching and Ray Data stage fusion implications

Target docs

• api-reference/processing-stage.mdx
• Ray Data/Xenna executor and infrastructure references
• Per-stage runtime/configuration cross-links
• 26.06 migration/release notes

Acceptance criteria

• Copy-pasteable Ray Data fixed-size and autoscaling examples are included
• Copy-pasteable Xenna override example is included
• Precedence, warnings, and invalid combinations are tabulated
• with_() examples cover resources, runtime env, worker counts, and stage specs
• Video fusion is explained as a resource-compatibility optimization, not a guaranteed behavior
• All enum/key names and defaults match merged code
• Fern checks pass

Related PRs: #2077, #2086, #2110