[BugFix] Make FlashInferMetadataBuilder non-blocking

[nvjullin]

Purpose

The blocking H2D memcpys breaks overlap scheduler #23569, setting them to non-blocking fixes it.
The correctness is ensured by vllm/v1/worker/gpu_model_runner.py:2112

            if self.prepare_inputs_event is not None:
                # Ensure prior step has finished with reused CPU tensors.
                self.prepare_inputs_event.synchronize()
            try:
                # Prepare the decoder inputs.
                (attn_metadata, logits_indices, spec_decode_metadata,
                 num_scheduled_tokens_np, spec_decode_common_attn_metadata,
                 max_query_len, ubatch_slices, num_tokens_after_padding
                 ) = self._prepare_inputs(scheduler_output)

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[nvpohanh]

@benchislett Please check if this fix is correct. Thanks!

[mgoin]

Looks reasonable to me, but @LucasWilkinson or @benchislett should validate before merge

[LucasWilkinson]

LGTM! Thanks for the contribution!