fix(tests/pytest_random_plugin.py): also seed Python's random

[yurekami]

Closes #11.

What

Add random.seed(seed) next to torch.manual_seed(seed) in the autouse
seed fixture so the --seed CLI option reproduces test inputs that
flow through Python's stdlib random.

Why

tile_kernels/testing/generator.py:94 (generate_rand_float) picks the
magnitude exponent with random.randint(-110, 126). The plugin only
seeds torch, so this exponent — and therefore the input scale used by
tests like tests/quant/test_per_block_cast_lossless.py — varies
across runs even when --seed is passed.

Reproduction

The bug is in a pure-Python entropy source, so it reproduces without a
GPU. Mirroring the fixture and generator.py:94:

run 1 exponents: [97, -60, 98, -69, 102, -84, 119, 2]
run 2 exponents: [80, -68, -47, 94,   0,  33, -74, 77]
run 3 exponents: [-69, -29, -106, 39, 87, 97, 19, 92]

Three back-to-back invocations with the same simulated --seed=0 and
the same node id produce three different sequences. With the patch,
both invocations within one process produce identical sequences.

Scope

• One import + one line. No behavior change for tests that don't touch
  Python's random.
• Does not touch NumPy — grep -r 'np\.random\|numpy.*random' is
  empty in this repo.