Fix Issue #995

[arn97]

Fixes a failure in GaussAdjoint when p === nothing.

Previously, the GaussAdjoint path assumed tunables supports length and Functors traversal. When p === nothing, this resulted in errors (e.g. length(::Nothing) or recursive traversal failures), causing Zygote.gradient(..., nothing) to error.

The fix represents the no-parameter case using an empty tunables container (Float64[]) while preserving repack = _ -> nothing to maintain RHS semantics. This allows Zygote.gradient(..., nothing) to correctly return (nothing,), consistent with InterpolatingAdjoint.

Checklist

• Appropriate tests were added: No new tests were added; verified with a minimal reproducer.
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated: Not applicable; internal fix
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

Testing:

1. Ran Pkg.test() locally.
2. Confirmed the reproduction snippet from Issue Nothing handling for GaussAdjoint #995 now succeeds: Zygote(loss2, nothing) returns nothing as expected.

Docs:
No documentation updates were required (internal fix; no public API changes).

I have not yet added a dedicated regression test; I can add one if you can point me to the preferred test location/pattern for SciMLSensitivity.

[ChrisRackauckas]

this isn't necessarily typed correctly, and would result in doing unnecessary computations