chore: Migrate Pixi system-requirements to custom platforms

[matthewfeickert]

• In Pixi v0.71.0 the system-requirements table was deprecated for custom platforms in the main workspace table. Add a 'linux-64-cuda' platform to support features with CUDA.
  • c.f. https://pixi.prefix.dev/v0.71.3/workspace/system_requirements/
• Update requires-pixi to 0.71.0 to ensure rich platform support.
• Note in the docs that users now need to remove platforms that do not support the custom platform's virtual packages, or that the custom platform needs to be explicitly given in the 'pixi add' command.

[matthewfeickert]

Note though that the semantics are now that one would need to run

pixi run --platform linux-64-cuda train

to have the task be found. pixi task add has --default-environment but it doesn't have --default-platform.

I'm not sure if this would be reasonable to request or not.

[matthewfeickert]

For

templates-GPUs/pixi/README.md

Line 65 in c954ba6

condor_submit mnist_gpu.sub

the mnist_gpu_<id>_0.err.txt logs contain the output of

templates-GPUs/pixi/mnist_gpu.sh

Line 21 in c954ba6

time pixi run python ./torch_detect_GPU.py

...
+ pixi run python ./torch_detect_GPU.py
 WARN Environment 'default' was resolved for a richer platform than 'linux-64' provides; this machine only meets the installed packages' minimum requirements, so it runs here by accident.

which shows that a Pixi is now resolving virtual packages from the machine doing the resolution (my laptop)

$ head -n 15 ./pixi.lock 
version: 7
platforms:
- name: p1
  subdir: linux-64
  virtual-packages:
  - __cuda=12.9
  - __unix=0=0
  - __linux=4.18
  - __glibc=2.28
  - __archspec=0=x86_64
environments:
  default:
    channels:
    - url: https://conda.anaconda.org/conda-forge/
    packages:
...

it might be necessary to make these virtual packages be more broad.

[matthewfeickert]

...
+ pixi run python ./torch_detect_GPU.py
 WARN Environment 'default' was resolved for a richer platform than 'linux-64' provides; this machine only meets the installed packages' minimum requirements, so it runs here by accident.

seems to be due to it landing on CHTC jobs, from

templates-GPUs/pixi/mnist_gpu.sub

Lines 19 to 41 in c954ba6

	# Require a machine with a modern version of the CUDA driver
	requirements = (GPUs_DriverVersion >= 12.0)
	# We must request 1 CPU in addition to 1 GPU
	request_cpus = 1
	request_gpus = 1
	request_memory = 10GB
	request_disk = 30GB
	+WantGPULab = true
	+GPUJobLength = "short"
	# Specify the GPU hardware architecture required
	# Check against the CUDA GPU Compute Capability for your software
	# e.g. python -c "import torch; print(torch.cuda.get_arch_list())"
	# The listed 'sm_xy' values show the x.y gpu capability supported
	#
	# Note that given
	# https://github.com/conda-forge/cudnn-feedstock/issues/124
	# there can be segfaults for cudnn>=9.11 on pre-Turing devices (<=sm_70)
	# so use sm_70 as a safer lower bound
	gpus_minimum_capability = 7.0

with

   Virtual packages: __unix=0=0
                   : __linux=5.14.0=0
                   : __glibc=2.34=0
                   : __cuda=12.8=0
                   : __archspec=1=cascadelake

when the environment is resolved to

  virtual-packages:
  - __cuda=12.9
  - __unix=0=0
  - __linux=4.18
  - __glibc=2.28
  - __archspec=0=x86_64

I think this either needs to have the submission areas tightened up to have more modern setups on CHTC or this should loosen the CUDA requirement to 12, which should be fine given how CUDA API/ABI works.