[xla:gpu] Refactor ThunkEmitter and collective (group) lowering

[ezhulenev]

NFC: this change doesn't haver any functional changes, and mostly tries to improve the structure and documentation on ThunkEmitter.

At run time XLA:GPU has one generic Async(Start/Done)Thunk that wraps nested synchronous compute. ThunkEmitter had async wrapping logic scattered around many places. This change makes thunk emission consistent with the runtime mental model and documents and cleans up ThunkEmitter.

• Separates HLO dispatch from operation-specific thunk construction.
• Centralizes AsyncExecution registration and sharing between start/done thunks.
• Adds explicit _collectives_group lowering for grouped collectives and device send/recv operations.
• Updates the send/recv combiner to create named, explicitly annotated async groups.
• Supports collective groups containing multiple communicators and degenerate device-to-device copies.

Motivation

The previous emitter mixed HLO dispatch, async lifecycle management, buffer selection, and thunk construction across several specialized paths. This made the semantics of generic async operations, legacy collectives, send/recv, host transfers, and grouped collectives difficult to reason about consistently.

The refactor makes those decisions explicit and ensures that:

• Async start and done thunks share the same registered execution state.
• Collective source and destination buffers follow HLO dataflow semantics.
• Device send/recv operations inside an explicit group are launched together.
• Host transfers continue to use their dedicated completion events.
• Synchronous and asynchronous collective lowering share the same payload emission logic.

[github-actions]

PR #46188 additions: 1124, deletions: 1036, total: 2160

🔴 This PR has a very large delta of over 1000. In order to enable an effective code reivew, please break the PR down into smaller and more focused PRs.
See the Small CLs Google Eng practice for more details on how to write compact PRs.

[github-actions]

PR #46188 additions: 1148, deletions: 1063, total: 2211

🔴 This PR has a very large delta of over 1000. In order to enable an effective code reivew, please break the PR down into smaller and more focused PRs.
See the Small CLs Google Eng practice for more details on how to write compact PRs.

[github-actions]

PR #46188 additions: 1150, deletions: 1063, total: 2213

🔴 This PR has a very large delta of over 1000. In order to enable an effective code reivew, please break the PR down into smaller and more focused PRs.
See the Small CLs Google Eng practice for more details on how to write compact PRs.

[github-actions]

PR #46188 additions: 1155, deletions: 1069, total: 2224

🔴 This PR has a very large delta of over 1000. In order to enable an effective code reivew, please break the PR down into smaller and more focused PRs.
See the Small CLs Google Eng practice for more details on how to write compact PRs.

[github-actions]

PR #46188 additions: 1163, deletions: 1073, total: 2236

🔴 This PR has a very large delta of over 1000. In order to enable an effective code reivew, please break the PR down into smaller and more focused PRs.
See the Small CLs Google Eng practice for more details on how to write compact PRs.