[FEATURE] Add profiling support for continuous batching

[nathon-lee]

Context

HybridEngineRollout supports optional profiling through
HybridEngineRolloutConfig.enable_profiling.

When continuous batching is enabled through
SamplingConfig.continuous_batch_size, the continuous path currently bypasses
the profiling logic. As a result, generate() completes without recording a
profiling snapshot, and get_last_profile() does not provide useful timing
information for continuous-batching workloads.

Proposed work

Define and implement profiling support for continuous batching.

The profiling contract should clarify how to measure and report:

• total rollout time;
• prompt prefill time;
• decode time;
• scheduler and cache-management overhead;
• number of decode forwards;
• generated tokens and throughput;
• active batch size and continuous-batching capacity.

Acceptance criteria

• enable_profiling=True produces a profiling snapshot for continuous
  batching.
• The reported fields have documented meanings for variable active batch sizes
  and independently retired rows.
• Profiling remains disabled by default.
• Enabling profiling does not change generated outputs or scheduling behavior.
• Tests cover the profiling output for a deterministic continuous-batching
  workload.

Follow-up from DeepSpeed PR #8368. cc @delock