pipeline: make pipeline_barrier context-aware, give each stage its own barrier

[aryanputta]

What this does

First concrete step toward the pipeline restructuring tracked in #100.

pipeline_barrier ignored its context parameter and always wrote events to the module-level _main_barrier_context singleton. All four barrier registrations in register_processing_functions shared one queue. That made independent per-stage barriers impossible and made the context parameter misleading — every call site passed event_pipe._main_barrier_context but the function never used it.

Root cause

# before
def pipeline_barrier(event: TraceEvent, _: AbstractContext) -> list[TraceEvent]:
    bctx = _main_barrier_context   # always the global, ignoring the argument
    bctx.collect(event)
    return []

Changes

src/aiu_trace_analyzer/pipeline/barrier.py

• _BarrierContext renamed to BarrierContext (public) so call sites can create instances without importing a private name.
• pipeline_barrier now uses ctx.collect(event) — the context it is passed — instead of the hardwired global.
• _main_barrier_context is kept as a BarrierContext() singleton for backwards compatibility; any call site that still passes it gets the same shared-barrier semantics as before.

src/aiu_trace_analyzer/pipeline/__init__.py

• BarrierContext exported so acelyzer.py and future PipelineStage code can create instances via the normal event_pipe namespace.

src/aiu_trace_analyzer/core/acelyzer.py

• Each of the four logical barrier groups in register_processing_functions now owns a named BarrierContext instance:

Name	Logical stage
normalize_barrier_ctx	frequency_align / normalize phase
overlap_barrier_ctx	overlap_tid detection (conditional)
comm_barrier_ctx	comm_summarize / rcu_util
categorize_barrier_ctx	event categorizer

No change to registration order or conditional logic.

tests/aiu_trace_analyzer/pipeline/test_barrier.py

• Updated existing test to pass a BarrierContext instance (previously passed None, which only worked because the function ignored its argument).
• test_independent_barrier_contexts_do_not_share_events: two barriers with separate contexts hold and drain their own events independently.
• test_barrier_drain_clears_hold: drain empties the hold list.

Validation

PYTHONPATH=src python3 -m pytest
# 157 passed, 5 xfailed

What this unblocks

With barriers context-aware and BarrierContext public, a PipelineStage class (next step in #100) can own its barrier instance and call pipeline_barrier correctly without any global state. The _main_barrier_context singleton is still available for the shared-barrier case but is no longer required.

[aryanputta]

Step 2 building on this is in #102 (PipelineStage abstraction). Both are ready for review when you have a chance, @lasch.

[aryanputta]

Superseded by #102 which combines both steps into one PR.