Feature title:  Determine cubin vs PTX for Nvidia

Feature Overview:
There are several tradeoffs to be considered for distributing CUDA binaries. PTX is a lot safer, while having the tradeoff of having some startup time as it's compiled for the first time onto the target GPU. Cubin avoids that startup time, but has the downside of if there's any issues, they won't be able to be caught by the PTX compiler.

Additionally, one other benefit of using PTX is its forwards and backwards compatibility, which could likely improve our build times.

Product(s) associated:

RHAIIS: y
RHEL AI: y
RHOAI: y

Goals:

• Determine which approach (or hybrid to use) of compilation of cubin vs PTX

Requirements:
An ADR documenting any changes that need to be made

Done - Acceptance Criteria:
A consensus on what to do, and an ADR to document it

Documentation Considerations :
We'll want to document this as an ADR, and any build changes that occur as a result will likely need to be documented as well