Let a recipe train at the precision its upstream reference uses

[EHxuban11]

RF100-VL forced every family to fp32. For EdgeCrafter that matches neither its
authors nor our own library, so the ec recipe now runs fp16, and the harness
records which precision produced a result.

Why ec is fp16

• Upstream ecdetseg/configs/ecdet/ecdet.yml sets use_amp: True with a
  GradScaler, and the train loop wraps the forward in torch.autocast with
  no dtype, which is fp16 on CUDA. All four size configs __include__ it, so
  fine-tuning inherits it too.
• LibreYOLO's own ECConfig sets amp: bool = True and overrides no
  amp_dtype, so it inherits the base "float16".
• LibreYOLO enables the GradScaler for fp16 and skips it for bf16
  (trainer.py:1599), so fp16 gets the loss scaling it needs.

Three bugs this also fixes

amp_dtype ignored the recipe. It was "amp": precision == "bfloat16" with
a hardcoded "amp_dtype": "bfloat16", so an fp16 recipe would have trained in
fp32 while its stats and submission claimed fp16. Both now come from one
PRECISION_AMP_DTYPE table.

Two separate gates rejected fp16. load_recipe refused it, and
_recipe_repro independently flagged it as mismatched training metadata,
which would have invalidated a submission after a campaign had already run.

No submission recorded its precision. config omits it and
runtime.precision is assemble_result's default, never passed by the
RF100-VL path, so every published run says fp32 whatever it did. There is now
rf100vl.training_precision, read from the per-dataset stats, and checkpoints
that disagree produce an invalid_reason instead of passing unnoticed.

Tests

162 passed. New coverage: the dtype table across all three precisions
(the regression above), the ec recipe pinned to fp16, and the existing
all-families check turned into an allowlist so a new family cannot pick a
non-fp32 precision without naming a reason.

Code provenance

Original code written for this PR by me, an AI agent. Precision facts were
read from two sources and are cited above: the EdgeCrafter reference
implementation at Intellindust-AI-Lab/EdgeCrafter (Apache-2.0), inspected
only to determine what precision it trains at, with no code copied or adapted;
and LibreYOLO's own training/config.py and training/trainer.py.

Unrelated finding while tracing that: THIRD_PARTY_NOTICES.txt in the
libreyolo repo attributes EdgeCrafter to github.com/EC-codehub/EdgeCrafter,
which 404s. The live repo appears to be Intellindust-AI-Lab/EdgeCrafter.
Worth correcting there, separately from this PR.