Exclude trailing padding token from the fine-tuning loss

[yentur]

Problem

default_loss counts one padding position per sequence. iterate_batches right pads
each row with zeros and reports lengths[:, 1] = L, the number of real tokens, but the
mask selects L - offset + 1 targets and the extra one is the pad value 0.

So every fine-tuned sequence is trained to predict token id 0 after its last token, and
ntoks is one too high. With --mask-prompt that padded position is a large share of
the signal: a six token prompt with a three token answer counts four targets per row
instead of three.

Root cause

mlx_lm/tuner/trainer.py:92

steps = mx.arange(1, targets.shape[1] + 1)
mask = mx.logical_and(steps >= lengths[:, 0:1], steps <= lengths[:, 1:])

targets[k] == batch[k + 1], so steps[k] is the index of targets[k] in the
un-shifted row. Real tokens occupy 0..L-1, so the last valid target index is L - 1
and steps <= L reaches one past the end.

Fix

steps <= lengths[:, 1:] becomes steps < lengths[:, 1:].

step == L is in range only when targets.shape[1] >= L, which happens only if the row
was padded, so this cannot drop a real target. Rows truncated to max_seq_length have
no padding column and come out unchanged. Reported train and validation loss will move
slightly, since ntoks is the denominator.

Verification

Four rows, six token prompt and three token answer, driven through the real
iterate_batches and default_loss. Before:

ntoks counted by default_loss : 16
real completion targets       : 12

loss with padding filled by 0  : 4.781232
loss with padding filled by 77 : 4.754244

After:

ntoks counted by default_loss : 12
real completion targets       : 12

loss with padding filled by 0  : 4.745682
loss with padding filled by 77 : 4.745682

The second pair is the invariant worth holding: changing only the bytes in the padding
region must not change the loss.

The two new tests fail on main,

$ python -m pytest tests/test_finetune.py -k TestDefaultLoss -q
E       AssertionError: 8 != 6
E       AssertionError: 3.734814167022705 != 3.6676673889160156
2 failed, 15 deselected

and pass with the fix:

2 passed, 15 deselected in 6.35s

test_finetune.py, test_tuner_trainer.py, test_tuner_utils.py and test_losses.py
give 29 passed. test_datsets.py::TestDatasets::test_hf fails on an unmodified checkout
too, huggingface_hub rejects the bare billsum name.

A 20 iteration LoRA run on mlx-community/Qwen2.5-0.5B-Instruct-4bit with
--mask-prompt trains normally, val 2.249 to 0.585.