feat(tflite): decode segmentation masks in TFLite inference - #1053

PILImage.Resampling was introduced in Pillow 9.1; older environments raise
AttributeError. Replace with a module-level fallback that prefers
Resampling.BILINEAR when available and falls back to the legacy
PILImage.BILINEAR constant.

[resolve roboflow#1] @Copilot / @Borda (gh):
roboflow#1053 (comment)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…didates

Previously the shape-based fallback called `next(...)` over all rank-4 outputs,
silently picking the first one. If a future model variant or onnx2tf version
emits an auxiliary rank-4 tensor, the wrong tensor would be decoded as masks
without any diagnostic.

Now the fallback collects all rank-4 candidates:
- exactly 1 → accepted as the mask output (same as before)
- ≥ 2 → log a warning and skip mask decode (fail safe, not silent corruption)

The name-based path (`"masks" in output name`) is unchanged and takes priority.

Resolve /review finding roboflow#1 (foundry:sw-engineer HIGH)
from .reports/review/2026-05-19T10-06-08Z/review-report.md

---
Co-authored-by: Claude Code <noreply@anthropic.com>

When keep.any() is False (all scores below threshold), mask_idx may still be
set for a segmentation model, causing _decode_masks to receive a (0, Hm, Wm)
input and returning a (0, H, W) bool array stored as Detections.mask. Callers
checking `dets.mask is None` would behave differently from `len(dets) == 0`.

Now mask decode is gated on `keep.any()`, so segmentation models return
`Detections.mask = None` when zero detections survive — consistent with the
detection-only model behaviour.

Resolve /review finding roboflow#6 (foundry:sw-engineer MEDIUM)
from .reports/review/2026-05-19T10-06-08Z/review-report.md

---
Co-authored-by: Claude Code <noreply@anthropic.com>

… edge cases

Resolve /review findings roboflow#4, roboflow#5, roboflow#8, roboflow#9, roboflow#10 from
.reports/review/2026-05-19T10-06-08Z/review-report.md

- test_run_inference_name_based_mask_detection: exercises primary name-based
  path (`"masks" in output name`) — previously dead code under test; every
  existing seg test used `Identity_N` which bypassed it
- test_run_inference_seg_model_no_detections_returns_none_mask: seg model +
  all logits below threshold → `Detections.mask is None` (post-keep.any() gate)
- test_decode_masks_raises_on_wrong_rank: _decode_masks rejects rank-4 input
  with a clear ValueError (exercises new guard from prior commit)
- test_decode_masks_exact_zero_logit_decodes_to_false: boundary at strict >0;
  uniform-zero logit map → all-False (not all-True)
- test_decode_masks_non_square_logit_input: (3,7,14) logits → (3,28,56) output,
  guards PIL bilinear resize for non-square feature maps
- test_decode_masks_parity_positive_negative_regions: high-magnitude ±10 logits
  verify correct True/False regions after bilinear upsample; parity regression
  guard for PostProcess.forward equivalent contract

---
Co-authored-by: Claude Code <noreply@anthropic.com>