feat(wd): drive inference from a profile sent by the caller

yu-server owns the tagger profile registry and its parser. This teaches the
sidecar to accept the inference-relevant part of a profile on the wire, so a
second family (camie) works without a second hardcoded recipe here — and
without a second copy of the profile parser in a second language.

Because half of that already existed and was unsafe. `load_input_size` and
`load_category_map_from_profile` picked "the first .json in the model
directory": `read_dir` order is unspecified, and for camie the first .json is
the 7 MB tag metadata, not a profile. Both are deleted here.

`profile` — `model_file`, `preprocess_spec`, `tag_source`, `output_spec`,
`default_thresholds`, `categories_mode`, `supports_categories`, deserialized
verbatim under their Python names. Omitted means the built-in WD recipe, which
is what every caller predating this field expects; that recipe is now
expressed as a profile too, so there is one code path rather than two.

`default_thresholds` is the *resolved* table. Whether the user's configured
thresholds or the profile's defaults win differs per adapter family, and that
decision belongs with the registry, not here.

The response carries `profile_applied`, so a caller can tell an old sidecar —
which would ignore the field and silently run the WD recipe against a camie
model — from one that honoured it.

A recipe this build cannot reproduce (`resize_strategy`, `channel_order`,
`layout`, `activation`, `tag_source.type`) is an error, and an `output_key`
naming a head the model does not export is an error. Approximating either
returns confident, well-formed, wrong tags; nothing downstream could detect it.

File names arriving in the profile are checked to be plain basenames before
being joined onto the model directory.

- **alpha** — the old code did `to_rgb8()`, which discards the alpha channel.
  Python composites over the pad colour first, so a transparent PNG resolved
  to white there and to whatever hid behind alpha=0 here.
- **scaled dimensions** — `DynamicImage::resize` rounds; Python's
  `int(w * scale)` truncates. A one-pixel difference shifts the paste offset.
- **unmapped CSV categories** — Python falls through to the raw value; this
  defaulted to "general", which would mislabel a whole category the profile
  forgot to map.
- **ragged CSV rows** — Python's DictReader tolerates them; the csv crate
  rejected them, so one side loaded the vocabulary and the other errored.
- **duplicate JSON keys** — Python raises; serde_json kept the last value.

The cached engine holds the vocabulary, thresholds and preprocess recipe, so
the model id alone no longer identifies it. The key now includes a fingerprint
of the resolved profile; the maps inside it are ordered so the fingerprint is
stable.

- **Against the real 752 MB camie export**, compared to tags produced by the
  Python adapter on the same three fixtures: **identical tag sets, categories
  and ratings** (16 / 23 / 20 tags), worst confidence delta **0.00005** — the
  granularity of Python's `round(conf, 4)`. `tests/camie_parity.rs`, skipped
  unless the model path is supplied.
- **Fault injection: 12 attempts, 11 caught.** Wrong output head, skipped
  sigmoid, ignored channel order (both directions), ignored NCHW, rounded
  instead of truncated, dropped mean, dropped scale, mean applied before
  scale, lexicographic tag ordering, fixed threshold instead of the
  per-category table, dropped alpha composite. The twelfth attempt failed to
  compile and is recorded as **not** a detection.
- The tensor builder is split out so channel order, normalisation order and
  layout are checked without the model — the camie fixtures are RGB/NCHW and
  cannot discriminate a build that ignores those fields.
- `cargo test -p yu-hailo-infer-core`: 54 passed. Workspace serially: 65
  passed. Two tests fail under parallel execution
  (`panic_discards_the_active_handle`, `speech2text_transcribe_rejects_invalid_task`);
  both pass serially and neither touches WD — a pre-existing isolation issue.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: eauesque