build(data_collection): make TORCH_CUDA_ARCH_LIST a --build-arg (unblocks Blackwell / sm_120)

[InfraestruturaRobotec]

What

Promote TORCH_CUDA_ARCH_LIST in source/data_collection/dockerfile from a
hard-coded ENV to an ARG with the same default, so it can be overridden at
build time.

ARG TORCH_CUDA_ARCH_LIST="8.9"
ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}

No behaviour change for anyone who does not pass the flag — the default is
byte-for-byte what it was.

Why

The pin sits immediately before cuRobo is built, so the resulting image can only
run cuRobo on Ada. On Blackwell (RTX 50-series, sm_120) every kinematics call
fails at the first launch:

CUDA error: no kernel image is available for execution on the device
  curobo/curobolib/kinematics.py:73 -> kinematics_fused_cu.forward(...)

Confirmed with cuobjdump, the shipped extensions contain sm_89 only:

kinematics_fused_cu.cpython-311-x86_64-linux-gnu.so -> sm_89

With this change, a Blackwell user builds with:

docker build --build-arg TORCH_CUDA_ARCH_LIST="8.9;12.0+PTX" \
  -f source/data_collection/dockerfile source/data_collection

and all five cuRobo extensions carry sm_120.

Tested on

RTX 5090 (sm_120), driver 580.126.09, torch 2.7.0+cu128, Isaac Sim 5.1,
cuRobo v0.7.6. After rebuilding with the override, MotionGen.plan_single
succeeds and a full geniesim autocollect data-collection run completes with
zero CUDA errors.

Note for 50-series users beyond this PR

This change is necessary but not sufficient on sm_120. Two further things are
needed, and neither belongs in this PR:

1. --no-deps when reinstalling cuRobo, otherwise pip re-resolves its
   dependency tree and replaces Isaac Sim's bundled torch 2.7.0+cu128 with a
   CUDA-13 stack, breaking the simulator.
2. -maxrregcount=64 in cuRobo's own nvcc args — the L-BFGS kernel then hits
   too many resources requested for launch because sm_120 codegen exceeds
   the 64K registers-per-block limit at threadsPerBlock = v_dim. Filed
   upstream as L-BFGS kernel: 'too many resources requested for launch' on Blackwell (sm_120) — register pressure, not shared memory NVlabs/curobo#716.

Happy to add a short README note pointing 50-series users at those if that would
be useful.