perf(api_core): avoid full-corpus scans in list_funcs/func_query/entity_query pagination

[lich0821]

Problem

list_funcs, func_query, and entity_query all built the entire
function/global/import corpus into a Python list before filtering and
paginating it — even for a bounded page (e.g. count=50). On binaries with
hundreds of thousands of functions this dominates per-call latency, and
_collect_entities("functions", ...) always paid for a per-row segment-name
lookup and a type-info probe even when the caller's requested fields never
use them.

Fix

• Add lazy_paginate_filter (utils.py): walks the underlying iterator
  lazily and stops as soon as enough matching rows are collected for the
  requested page, instead of materializing every row up front. Falls back to
  full materialization only when count=0 (unbounded page), since computing
  next_offset correctly still requires the full filtered set in that case.
• _collect_entities takes needs_segment/needs_has_type flags so
  entity_query can skip the segment-name lookup and type-info probe per
  function when neither the active filter nor the requested output fields
  need them.
• list_funcs/func_query switched to the lazy path for bounded pages.

Existing callers and output shape are unchanged; this only changes how many
rows get built before filtering stops.

Test plan

• Existing test_api_core.py pagination/filter assertions pass unchanged
  (ran via ida-mcp-test tests/crackme03.elf / typed_fixture.elf)