ui: Heap Dump Explorer — multi-trace baseline diff mode

Adds baseline / current / diff workflow to the Heapdump Explorer plugin.
The user can pool one or more reference traces, pick a dump from any
pooled trace as the active baseline, and the Classes / Strings / Arrays
/ Bitmaps / Dominators / Overview tabs render side-by-side current vs
baseline values with classified delta status (NEW / REMOVED / GREW /
SHRANK / UNCHANGED).

Architecture:
* Each pooled baseline trace owns its own WasmEngineProxy (separate
  Web Worker) — a single sqlite ATTACH across workers is impossible,
  so each side runs its filter independently and the merge happens
  in JS (diff/diff_rows.ts). Outer-join over ~20K x 20K rows is
  cheap (~10ms).
* In-flight diff fetches snapshot the active primary + baseline refs
  before they await; if either ref differs after, the merge is
  skipped. Same check in catch silently abandons fetches whose
  engine was disposed mid-flight.
* The Overview cache keys on (engine identity, filter SQL); both
  primary and baseline engines need their own cached entry, and
  passing the primary's dumpFilterSql() to the baseline engine
  would filter on (upid, ts) values that exist in the primary trace
  but not the baseline (silent zero-row result).
* Tabs widget keys on (primary upid+ts, baseline trace+upid+ts);
  selection changes remount the diff view, the snapshot guard
  handles the tail of the previous run.
* Status classification comes from row presence, never from
  value==0 (a class with dominated_size=0 in baseline but present
  in both snapshots must be UNCHANGED, not NEW).
* Hprof has no real pid (trace_processor reports 0); the header
  and popup labels render "Java heap dump" instead of "pid 0" so
  the selector reads naturally for both pftrace and hprof inputs.

Change-Id: I2f69614c7d87091b3b301ff51e110b70e2738a40