worker/executor with subprocess executor

[khimaros]

workaround for #10997 and https://gitlab.freedesktop.org/drm/amd/-/issues/4632#note_3194291

• pull prompt execution out into a separate module
• decouple prompt execution from server with a strict interface/protocol
• add two new workers: native and process worker
• process worker starts a brand new GPU context for each prompt

the 6.18 kernel releases should resolve the underlying kernel bug, but i thought this might be useful for people in the meantime and i think it may be a useful architectural change generally.

for example, it could allow for networked workers on other machines.