Fix KITTI eval and checkpoint loading on numpy>=2 / torch>=2.6

[blancnight2]

What

Two independent compatibility fixes needed to run OpenPCDet on current toolchains
(numpy >= 2.0, torch >= 2.6). Both currently block evaluation out of the box.

1. rotate_iou.py — drop the numba dependency

The numba.cuda kernel used by the KITTI evaluator fails type inference on recent
numba releases combined with numpy >= 2.0, raising Signature mismatch from
FixupArgs. The workaround circulating in the issue tracker is to pin
numpy==1.23, which is not viable on newer toolchains (and conflicts with other
packages in a modern CUDA 12.x environment).

Instead of pinning numpy back, this PR removes the numba.cuda GPU-kernel path from KITTI evaluation (the part that breaks); the CPU @numba.jit functions in eval.py are untouched and continue to work by reimplementing rotate_iou_gpu_eval as a thin PyTorch
wrapper around the project's own compiled boxes_overlap_bev_gpu op
(pcdet/ops/iou3d_nms). That op is already a hard dependency of OpenPCDet, so
no new dependency is introduced — one is removed.

Geometry equivalence. The original code builds corners and rotates them by
[[cos, sin], [-sin, cos]], which is R(-angle), whereas boxes_overlap_bev_gpu
rotates by R(heading). Setting heading = -angle (with dx = x_d, dy = y_d)
makes the two rectangles identical, so the returned intersection area matches.

2. detector3d_template.py — weights_only=False

torch 2.6 flipped the default of torch.load to weights_only=True. OpenPCDet
checkpoints contain more than plain tensors, so loading them now fails. This PR
passes weights_only=False explicitly at the two checkpoint load sites.

Verification

rotate_iou_gpu_eval was checked against an independent shapely reference that
reimplements the original corner convention, over random boxes and all four
criterion values:

criterion	meaning	max abs error
-1	IoU (inter / union)	9.0e-07
0	inter / area(boxes)	1.3e-06
1	inter / area(query_boxes)	6.2e-07
other	raw intersection area	6.9e-05 (area units)

i.e. float32 round-off. End-to-end, PointPillars trained for 80 epochs evaluates
normally on KITTI val with this evaluator (3D AP R40 — Car 85.59 / 75.60 / 72.76,
Pedestrian 50.79 / 43.75 / 39.12, Cyclist 79.72 / 62.05 / 57.71).

Environment

Ubuntu 20.04 (WSL2), torch 2.11.0+cu128, CUDA 12.8, RTX 5070 (sm_120), numpy 2.2,
spconv-cu124.

Notes

• device_id is now honoured (tensors are placed on cuda:device_id) instead of
  being ignored.
• The deprecated torch.cuda.FloatTensor(...) constructor is avoided.
• The evaluator now requires the compiled iou3d_nms op at eval time. This is
  already required to run any detector in OpenPCDet, so it should not be a new
  constraint in practice — happy to guard it behind a try/except with a clear
  error message if maintainers prefer.

[blancnight2]

Two clarifications after re-reading my own description.

1. Scope of the numba change. To be precise, this removes the numba.cuda
GPU-kernel path from KITTI evaluation, not numba altogether. eval.py still uses
CPU @numba.jit(nopython=True) in five places and those are unaffected — only the
CUDA kernel hits the type-inference regression (Signature mismatch from
FixupArgs). Apologies for the overstatement in the description; I have corrected
it there as well.

For reference, the failure is limited to evaluation — training and inference never
touch this code path. The weights_only change in (2) is broader: it affects any
checkpoint load, including resuming training, test.py, and the demo scripts.

2. The same kernel is duplicated elsewhere. The identical numba.cuda
implementation of rotate_iou_gpu_eval also lives in
pcdet/datasets/once/once_eval/iou_utils.py (lines ~277-341), so ONCE evaluation
should fail the same way on numpy >= 2. I left it untouched because I have no ONCE
data to validate the change against, and I did not want to ship an unverified edit.
Happy to apply the same fix there if maintainers want it in this PR, or to leave it
for a separate one.

That makes three implementations of rotated-rectangle intersection in the repo:
the two numba copies above, and the compiled C++/CUDA boxes_overlap_bev_gpu in
pcdet/ops/iou3d_nms that this PR reuses.