Fix Day 1 GPU float32 attention test under default MLX precision

[skyzh]

Summary

• Compare the Week 1 Day 1 Task 1 unmasked float32 GPU attention output with a scoped 2**-11 absolute tolerance under MLX's default precision mode. This empirical margin covers the reported 1.1e-4 gap for the bounded fixture; it is not a hardware or general attention error bound. Scale and mask cases, CPU and float16 cases, and the shared comparator retain their strict checks.
• The test and test-refsol PDM scripts use MLX's default precision mode; neither sets MLX_ENABLE_TF32. Learner and reference implementations are unchanged.
• Keep the setup, preface, and Day 1 instructions that build both native extensions before the first tests.

Related to #322. This PR does not close the issue: the reported Apple M5 / MLX 0.32.1 case still needs a direct retest.

Evidence

• On M4 Pro / MLX 0.32.0 with MLX_ENABLE_TF32 unset, focused Day 1 reference tests pass 36/36. The full unfiltered reference suite on the executable-equivalent parent passes 679 with 2 optional skips. The unchanged learner starter collects, with 4 softmax passes and 32 expected attention TODO failures; a temporary implementation of the lesson formula passes all 36 Task 1 cases.
• Deliberately wrong scale, mask, batch, shape, and NaN outputs are rejected by the affected comparison. The scoped plain comparison accepts a +1.1e-4 offset and rejects +1e-3; scale and mask checks reject the smaller offset.
• Independent exact-local-head numerical, factual, learner, and static reviews passed on b6c8ea5465e1347d02827e16a3c523c6b5e78c3e. Hosted PR checks and public-head review are separate gates.

Limits

No direct Apple M5 / MLX 0.32.1 retest was available. The local synthetic rounding surrogate is diagnostic evidence, not a bound on MLX or M5. This PR makes no BF16 or model-performance claim.