fix(test-harness): unique --basetemp per parallel pytest subprocess

[Lexus2016]

Problem

scripts/run_tests_parallel.py runs one python -m pytest <file> subprocess
per file with bounded parallelism (up to cpu*2). None passed --basetemp,
so every subprocess used the shared default /tmp/pytest-of-<user>/ root.

pytest's numbered-basetemp retention GC ("keep the last few pytest-N dirs")
runs at session start and rmtrees sibling subprocesses' live basetemp
out from under them. Result: flaky
FileNotFoundError: /tmp/pytest-of-<user>/pytest-N at fixture setup
(notably the pytest_asyncio tmp_path fixture), with zero relation to the
code under test.

This is the harness flake that reddens unrelated PRs. Concretely, #1170's
slice 4/8 failed with ~126 tmp_path setup errors in
tests/gateway/test_telegram_noise_filter.py ("all tests passed but pytest
exited non-zero") that vanished on rerun.

Fix

Allocate a unique --basetemp (tempfile.mkdtemp) per subprocess, so the
shared parent — and therefore the cross-process rmtree race — no longer
exists. Details:

• An explicit user-supplied --basetemp in passthrough args is respected.
• The temp tree is reclaimed on every exit path (normal, timeout, interrupt),
  so local runs don't accumulate one tree per file per invocation. CI runners
  are ephemeral anyway.

Testing

Ran the real harness over the tmp_path-heavy files (including the previously
flaky one) in parallel:

scripts/run_tests.sh tests/gateway/test_telegram_noise_filter.py \
  tests/tools/test_delegate.py tests/tools/test_tool_search.py
=> 3 files, 426 tests passed, 0 failed

• ruff check clean; py_compile clean.
• Verified no hermes-pytest-* temp trees leak after the run (cleanup works).
• test_durations.json is gitignored (harness side-effect, not committed).

[Lexus2016]

Amended: the initial hermes-pytest- mkdtemp prefix put the literal "hermes" into the injected --basetemp path, which is passed as a subprocess arg inside some tests. conftest.py's live-system guard blocks any subprocess whose command contains "hermes"/"gateway" alongside a process-killer token — e.g. test_search_hidden_dirs.py greps for the literal "skill" (a token in _PROCESS_KILLERS), so the run reddened slice 4/8. Switched the prefix to a neutral pytest-rtp- so the injected path is behaviourally identical to pytest's default pytest-of-<user> location. Verified locally: the previously-failing test now passes through the harness, and the tmp_path-heavy files stay green.