perf: load brain plugins lazily, only when their target module is imported

[Pierre-Sassoulas]

Summary

astroid_manager eagerly loaded every brain plugin (~49 modules) at
import astroid so their transforms were registered up front. Most of
that work is wasted — brain_numpy_* does nothing for code that never
imports numpy, brain_pytest nothing without pytest, and so on.

• Data-driven brain registry (first commit, behavior-preserving
  refactor): replace the hand-maintained register_all_brains — a
  50-line import block plus 50 .register() calls — with an
  _EAGER_BRAINS tuple and a _LAZY_BRAINS mapping from a module-name
  trigger to the brain modules it needs. Brains load via importlib
  and register at most once per process.
• Lazy loading (second commit): register only the two universal
  brains (brain_builtin_inference and brain_type, which target
  import-less builtins) at startup, and load the rest on demand.
  AstroidBuilder._post_build scans a freshly built module's Import
  / ImportFrom nodes and AstroidManager.ast_from_module_name keys
  on the module being resolved; each calls load_brains_for_modname
  so a brain is registered before the transforms that need it run.

import astroid goes from 165 to 109 loaded modules (brain modules
51 → 2). Combined with #3062, 165 → 100.

test_recursion_error_trapped relied on the brain inference
predicates running on every Call node and exhausting the stack; with
lazy loading those predicates aren't registered, so it now force-loads
every brain to keep verifying the trap (trap behaviour unchanged).

Test plan

• Full test suite passes locally (1956 passed, 53 skipped, 16 xfailed)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.45%. Comparing base (2ef6d50) to head (56387a3).

Additional details and impacted files

@@                        Coverage Diff                        @@
##           codspeed-wizard-1774989768270    #3069      +/-   ##
=================================================================
- Coverage                          93.56%   93.45%   -0.11%     
=================================================================
  Files                                 92       92              
  Lines                              11345    11324      -21     
=================================================================
- Hits                               10615    10583      -32     
- Misses                               730      741      +11     

Flag	Coverage Δ
linux	93.32% <100.00%> (-0.11%)	⬇️
pypy	93.45% <100.00%> (-0.11%)	⬇️
windows	93.42% <100.00%> (-0.11%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
astroid/astroid_manager.py	100.00% <100.00%> (ø)
astroid/brain/helpers.py	100.00% <100.00%> (ø)
astroid/builder.py	96.21% <100.00%> (+0.09%)	⬆️
astroid/manager.py	90.24% <100.00%> (+0.07%)	⬆️

... and 13 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[codspeed-hq]

Merging this PR will degrade performance by 2.59%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

❌ 1 regressed benchmark
✅ 2 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	test_bench_endtoend_parse_flask	2.2 s	2.3 s	-2.59%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/lazy-brain-loading (9ee39c6) with main (c50a1f4)}