[xla:gpu] Add async thunk passes for optimizing async execution

[ezhulenev]

Thunk passes that correct LHS scheduling decisions if buffer assignment permits it:

• Implement RemoveRedundantAsyncThunkPass — removes redundant async start/done thunk pairs where an AsyncDoneThunk immediately follows its matching AsyncStartThunk. In this case there is no actual asynchronous execution (the done thunk just waits for an event recorded right before it), so we inline the nested thunk sequence from the start thunk, avoiding the overhead of creating an async execution scope (recording events and synchronizing streams).

• Implement ExpandAsyncScopeThunkPass — a new thunk pass that widens async execution scopes by moving AsyncStartThunk as far up the thunk sequence as possible and AsyncDoneThunk as far down as possible, maximizing the overlap window between async operations (e.g., collectives) and compute.

• Conflict detection uses BufferUse::ReadWriteSet and ResourceUse::ReadWriteSet, with buffer and resource uses collected transitively via the Thunk::Walk API. AsyncDoneThunk (which reports no uses of its own) inherits the uses of all matching AsyncStartThunks to handle pipelined async chains correctly.

[seantalts]

otherwise looks good to me

[seantalts]

Could you add a couple of tests (I think these might fail today)

1. Redundant pair nested inside a non-redundant AsyncStartThunk: An AsyncStartThunk whose nested sequence contains [start_inner, done_inner], where the outer start and done are not adjacent. The inner pair should still get inlined.
2. Same for ExpandAsyncScopeThunkPass: An AsyncStartThunk whose nested sequence contains [kernel, start, kernel, done, kernel]. The inner scope should get expanded.

Probably also good to have a test that combines the passes and makes sure they interact well, and also test multiple independent async pairs.

[ezhulenev]

Looks like when we already have thunk sequence it's too late to optimize it, it must be done earlier at HLO level. Abandoning it for now.