camprojection emits non-finite data at zero camera depth

[dvasquezlopez]

Specula found this at commit
88af9cc682d42e2e925daa0bae25f53277999f7b. The affected sensor.py is
byte-identical on current main commit
dcaa09431b84304d1de1e0c1200d981843104440.

camprojection guards camera-local depth below MJ_MINVAL by clamping it to
[-MJ_MINVAL, MJ_MINVAL]. Every value entering that branch is already inside
the interval, so the clamp changes nothing. Zero reaches the projection divide
unchanged, while other near-zero values retain too-small magnitudes.

Current sequence

T0: A camera-local point has zero or near-zero depth.
T1: The guard clamps that depth into an interval that already contains it.
T2: The unchanged depth reaches the projection divide.
T3: The projection becomes non-finite or ten times too large.

Proposed sequence

T0: A camera-local point has zero or near-zero depth.
T1: The guard sets its magnitude to MJ_MINVAL and preserves its sign.
T2: The projection divides by a bounded nonzero depth.
T3: The result remains finite and matches MuJoCo.

The exact public-path differential witness ran on CPU and CUDA. Ordinary depth
matches MuJoCo. At zero, MuJoCo returns finite
[-5.794112549695424e17, 240], while Warp returns [-inf, nan]. At
+/-1e-16, Warp has ten times MuJoCo's magnitude. The full suite and
multi-world batching were not run.

Replace the guarded denominator explicitly and add CPU/CUDA parity tests for
negative near-zero, zero, positive near-zero, and ordinary depth.

@thowell was detected as the most likely owner from current-source blame.
Please feel free to assign this to someone else if another owner is closer to
the code.