PR #634 set MLXTrainingConfig.max_grad_value = 1.0 (later 5.0) and at
config-resolution time silently zeroed out a user-supplied
max_grad_norm. The elementwise clip rotates the gradient per leaf,
which is mathematically different from clip_grad_norm and not what
HF/TRL users opt into when they pass max_grad_norm=1.0. Same dataset,
same seed, same LR converges to a different basin on MLX than on CUDA;
greedy generation collapses to gibberish even though loss descends.

Make max_grad_value opt-in only:
  * Default None (off). User-supplied max_grad_norm is honored by
    default, matching HF/TRL semantics on CUDA.
  * Explicit float > 0 keeps the existing low-memory clip path AND the
    existing "ignoring max_grad_norm" notice when both are set.

Add a config-level regression test pinning the default to None.

Refs: #662, unslothai/unsloth#5498.

PR #634 silently flipped MLX AdamW's bias_correction from the historical
MLX default of False to True (matching torch.optim.AdamW). For real
multi-epoch fine-tunes the two converge identically after ~10-20
warmup steps, but for short memorization runs the difference is large:
bias_correction=True shrinks the step-1 effective update by ~3x.

Empirical bisection on a Mac M1 CI runner (probes 12 + 14 of the
mlx-parity-probes workflow):
  * pre-#634 trainer (bias_correction=False), 7 steps:
      loss 10.55 -> 5.04 (bouncy), generates "Unsloth! ..."
  * HEAD + PR #663 only (bias_correction=True), 7 steps:
      loss 10.55 -> 0.17 (smooth), generates "5 lbs!"
  * HEAD + bias_correction=False (this PR), 7 steps:
      loss 10.55 -> 2.44 (bouncy), generates "Unsloth! ..."

The upstream MLX smoke test in unslothai/unsloth and every other
existing MLX fine-tune script implicitly relied on the bias_correction=
False default. Restoring it as the default fixes that contract.

Add `adam_bias_correction: bool = False` to MLXTrainingConfig so users
who want true HF/torch.AdamW parity can opt in explicitly. Plumb it
through both the adamw and adam construction paths.

Regression test pins the default to False.

…re parity"

This reverts commit 72a448b.

…act)

Reverts commit 18596f2. The original 72a448b adam_bias_correction
exposure was reverted on the (premature) conclusion that bc=True at
30 steps with seed=3407 produced a working memorization. Subsequent
parity probing (rounds E-G of mlx-parity-probes) showed:

  * the bc=True trainer converges to post_train_loss ~0 across all
    seeds tested (3407, 42, 999, 1337, 7777, 12345) and a wide LR
    band (5e-4 - 2e-3) -- training is healthy;
  * BUT the post-train greedy-decode test ("does the model emit
    'Unsloth!' from the prompt?") is non-monotonic across (steps,
    seed) under bc=True:
      seed=3407: 30 OK, 50 BAD, 60 BAD, 100 OK
      seed=42:   30 OK, 60 BAD
      seed=12345: 30 BAD, 40 BAD, 50 OK, 60 OK
      etc.
  * mlx-lm's native LoRA at the same iter counts barely converges
    (last_loss 3-5 with mlx-lm defaults), so it doesn't reach the
    over-memorized basin our trainer does -- the fragility is a
    side effect of fast/aggressive memorization on a tiny single-
    row fixture, not a trainer bug.

Given the smoke fixture is so brittle to (steps, seed), users will
benefit from being able to flip bias_correction back to its MLX-
ecosystem-native False (which is what mlx.optimizers.AdamW,
mlx_lm.lora and every existing MLX fine-tune script use) without
also having to fork the trainer.

This commit:
  - re-adds `adam_bias_correction: bool = False` to MLXTrainingConfig
  - plumbs it through both the adamw and adam construction paths
  - default = False (MLX framework default, mlx-lm default, pre-#634)
  - users who want torch.optim.AdamW parity can opt in explicitly.

Round L of mlx-parity-probes located the divergence boundary for
adam_bias_correction=False on the single-row LoRA smoke fixture:
  50 steps:  post_train_loss=5.06 (high but finite)
  100 steps: post_train_loss=NaN (catastrophic divergence)

So the field isn't just "slow vs fast" -- bc=False is dangerous at
long horizons on a tiny / fast-overfitting fixture. Update the
docstring so users opt in to False with eyes open instead of
treating it as a safer "MLX-ecosystem default" option.