Make IndexedSet bulk removal linear via one-pass rebuild

[LD-RW]

Fixes #439

difference_update() and intersection_update() looped discard() per item.
Every 384 scattered removals, _cull() triggers a full O(n) _compact(),
making bulk removal of k items O(n·k/384) — quadratic when k scales with n.
See #439 for the flame graph and profile.

Change

• Name the magic 384 threshold: _MAX_DEAD_INTERVALS (used in _cull and here)
• Add private _bulk_discard(): rebuilds item_list/item_index_map in one
  O(n) pass and clears dead_indices
• difference_update/intersection_update take the bulk path only when the
  removal count exceeds the threshold; small removals keep the existing
  incremental path unchanged
• symmetric_difference_update deliberately untouched: its single-pass loop
  toggles order-dependently when other contains duplicates, which a bulk
  path cannot reproduce

Behavior

Identical: both methods already materialized the removal snapshot up front
(self.intersection(...) / self.difference(...)) before discarding, so
element order of survivors, indexing, and contents are unchanged. Verified with
randomized differential trials against the previous behavior (order,
__getitem__, and index() all checked) in addition to the test suite.

Numbers (Intel Core i7-12700H, Python 3.14.6)

difference_update removing every other element, best of 3:

n	before	after	speedup
10,000	7.8 ms	1.7 ms	4.6x
20,000	24.6 ms	3.8 ms	6.5x
40,000	86.1 ms	7.7 ms	11x
80,000	317.3 ms	15.3 ms	21x
160,000	1,197.9 ms	31.2 ms	38x

Scaling is linear after the change (doubling n doubles time).

Tests

• test_bulk_difference_update / test_bulk_intersection_update: exercise the
  new path (1000 elements > threshold), assert order, indexing, and post-op
  mutation behavior
• test_small_difference_update_unchanged: pins the incremental path for
  small removals