fix(book): two chapter examples trained to NaN, on main, for three months

[noahgift]

fix(book): two chapter examples trained to NaN, on main, for three months

Chapter Examples Run failed this batch on

Epoch 50: loss=NaN
thread 'main' panicked at ch10_training.rs:72:
Training must reduce loss: NaN < 270.3525

and the same shape in ch24_switch_pytorch.rs:61.

NOT introduced here. Both fail identically on origin/main — verified by running
them in a clean worktree at main (ch10 rc=101 NaN < 338.4173, ch24 rc=101
Training must reduce loss). They have been broken since at most 2026-05-14,
which is the last time mdBook CI ran on main: the workflow filtered on
paths: book/**, and nobody had touched book/ since. Fourth instance of that
class in this one job, each revealed only by fixing the one before it.

ROOT CAUSE, and the distinction that matters: this is the EXAMPLE's learning
rate, not a framework defect. Both examples train on x up to 8.0 and y up to
15.0, unnormalized, at lr 0.01 — large enough first MSE gradients that the step
overshoots and the loss reaches NaN by epoch 25. Before changing anything I swept
the rate on the unmodified example:

lr 0.01   -> NaN                         (panic)
lr 0.001  -> 110.3778 -> 0.0000          (rc 0)
lr 0.0001 ->   1.0728 -> 0.2124          (rc 0)

SGD, MSELoss and backward are all correct at a step size the data supports. Had
the sweep NOT converged, this would have been a P0 library bug rather than a doc
fix, which is why it was worth establishing first.

Both examples now use lr 0.001.

Weight init is random per run, so the initial loss varies widely and a single
green run proves little. Five consecutive runs of each, 10/10 rc=0:

Initial: 159.1656 -> Final: 0.2184
Initial:  85.7478 -> Final: 0.0239
Initial:  32.6145 -> Final: 0.0007
Initial:  81.9896 -> Final: 0.0635
Initial: 159.7258 -> Final: 0.1001

Every run converges by more than two orders of magnitude, so the
final_loss < initial_loss assertion has a wide margin rather than sitting on a
knife edge — the failure mode that made an earlier golden gate a coin flip.

Refs #2373

---

Split from #2458 only because that PR had already entered the merge queue and
queued branches cannot be updated. These two files are independent of it, and
the examples are already broken on main, so #2458 is not a regression either
way.