fix: make generate_session_id independent of the global random seed

[pcbeingused333]

Problem

generate_session_id() picks its adjective and noun with random.choice:

adj = random.choice(adjectives)
noun = random.choice(nouns)

set_seed() reseeds Python's global random state. So any run that calls set_seed(N) and then generate_session_id() gets the same word pair every time — only the secrets.token_hex(2) suffix varies:

set_seed(1234); generate_session_id()  # 20260908_1607_frosty_morning_07da
set_seed(1234); generate_session_id()  # 20260908_1607_frosty_morning_8e1d

Since set_seed() is the normal way to use this reproducibility-focused framework, session ids collapse to near-identical strings across runs, which defeats the point of a human-readable per-run identifier.

Fix

Use secrets.choice for the words too (secrets is already imported for the hex suffix). set_seed() still reseeds random/numpy/torch exactly as before — only the session-id word choice is moved off the seeded RNG.

Tests

Added test_word_choice_is_independent_of_the_global_seed: seeding with a fixed value and calling generate_session_id() repeatedly still yields varying word parts. It fails on main (frosty_morning pinned by the seed) and passes with the change. pytest tests/unit/test_reproducibility.py is green (19 passed); ruff check / ruff format clean.

Note: the pre-commit check's ty hook is red on main too (commit 0d1862e, on fenn/experimental/vision/normalize.py and fenn/logging.py), unrelated to this change.

[pcbeingused333]

The pre-commit check is red on the ty hook, but that's pre-existing on main (commit 0d1862e, current HEAD, fails the same job on fenn/experimental/vision/normalize.py and fenn/logging.py — neither touched here). ruff check / ruff format pass; the four test jobs are the relevant ones and they're green.