[Performance] Optimize encoder cache memory consumption by storing encoder outputs only

[imkero]

Purpose

Fix #25903

This PR modifies model runners (they operates encoder cache) & multi modal Placeholder (they provide the metadata) to count and cache the original encoder cache instead of the interleaved encoder cache.

In detail, this PR includes following modifications:

1. add post-init behaviour in PlaceholderRange:
   • cache the num_embeds
   • remove the leading & tailing False in is_embed to enable more concrete scheduling
2. modify the encoder cache behaviour
   • cache the original encoder_output (instead of the interleaved one currently)
   • slice the cached encoder_output by counting True in PlaceholderRange.is_embed mask
3. modify Request.get_num_encoder_tokens to return actual encoder tokens num (without non embedding tokens) for encoder cache scheduling, coresponding to the modified encoder caching behaviour

Still in draft part

• modify the profiling behaviour?
• modify other model runners besides gpu_model_runner
• enhance the performance of counting True in is_embed

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.

[imkero]

I'm wondering we can produce an is_embed_cumsum ahead of time so that we do not need to count True in is_embed in every step