Port 1.7.3 and 1.7.4 fixes to dev - #1037

…cache (#989)

…989)

Release 1.7.2

onnx2tf's -oiqt builds the integer model twice: first with float32 I/O,
then again with int8 I/O, and only the second is shipped. The I/O dtype
does not change the calibrated scales, so the first pass is overhead.

Build the SavedModel with onnx2tf and run a single in-process
TFLiteConverter pass instead. On a YOLO9-t fine-tune at 800x1280 with 156
calibration images this takes 11.6 min instead of 74.2 min, with mAP50-95
0.621 vs 0.625 and identical quantization scales.

Calibration still feeds one image per sample: the converter requires
every sample to match the model input shape, so batching is not possible
for a batch-1 graph.

int8_backend='onnx2tf' keeps the previous CLI route, and
int8_output_dtype='float32' writes float outputs for host-dequantizing
runtimes.

Quantize TFLite INT8 once instead of twice

- huggingface_hub is a core dependency (>=0.34, below the molmo2 pin)
- the weight selector prefers the checkpoint named after the repo, so the
  YOLO9 and YOLOX repos with a legacy lowercase copy load by repo id
- a differently cased libreyolo/ folder no longer shadows Hub ids on
  case-insensitive filesystems

…oint

#1003 set TensorFlow's private converter attribute by name, which the
repository vocabulary test reads as a capability label. Same split as the
test's own word list; behaviour unchanged.

Release 1.7.3: load models by Hugging Face repo id on a plain install

- Environment markers skip backend deps that have no wheels for the
  running Python (onnx2tf, executorch/torch/torchvision, tensorrt-cu12,
  coremltools on 3.14; paddlepaddle, x2paddle and pins on 3.13+), so
  pip no longer backtracks to an old libreyolo to satisfy an extra.
- Backends and exporters raise a clear error naming the Python they need.
- Install hints are quoted: pip install "libreyolo[x]".
- README: Hub one-liner, CPU-only torch note, absolute links.
- Dependency resolution check covers Python 3.14.

- TensorRT export, INT8 calibration and inference say that the pinned
  tensorrt-cu12 has no Python 3.14 bindings and point to 3.13 or earlier.
- The CoreML platform-hint test pins Python 3.12 so it passes on 3.14,
  and a separate test covers the 3.14 message.

Release 1.7.4: extras no longer downgrade LibreYOLO on new Pythons

Brings the 1.7.2-1.7.4 release commits (#992, #1003, #1020, #1026) into dev.

- pyproject: keep dev's version line, bumped to 1.8.0.dev0 so dev sorts
  above the released 1.7.4; take the core huggingface_hub dependency and
  the Python 3.13/3.14 extras markers.
- CHANGELOG: dev's own entries stay under [Unreleased]; the #989 entries
  that dev had listed there move to the released [1.7.2] section, followed
  by release's [1.7.3] and [1.7.4] sections.
- README.md, README.zh-CN.md, SPONSORS.md: dev's versions (#1033, #1035).

hf_hub.py: keep the repo-named checkpoint pick from 1.7.3, then dev's
sorted error with an example call. docs/hf_hub.md: keep both bullets.

…e the NumPy source-build exception to vla