Update pixi.toml to latest MAX release, add support for aarch64 Linux

[BradLarson]

The latest nightly version of MAX adds full support for running the GPU puzzles on DGX Spark and Jetson devices. These systems are aarch64 Linux hosts, so we also need to add linux-aarch64 entries. I believe PyTorch only started supporting aarch64 for CUDA in 2.9.0, so that version is used for the imports on this architecture. PyTorch properly installs and tests pass on DGX Spark at that version, and not for the older import used in linux-64.

All tests pass on DGX Spark except for the last three, which will need upstream kernel enhancements to reflect this new architecture. They should be supported on this device, but we haven't yet added the appropriate checks yet.

Additionally, there are some significant Mojo API changes that this nightly starts to support. The migration to those will come in a follow-on PR to keep this one focused.

[BradLarson]

Hmm, we need to set MODULAR_NVPTX_COMPILER_PATH for the CI system to work correctly our current nightlies, but the CI systems don't appear to have ptxas installed or at least not installed in the normal location. More investigation may be needed.