Add RTX-focused fast path preset via config profile

[vecnode]

Add an opt‑in “fast path” configuration preset tuned for typical RTX GPUs that reuses the existing configuration system but applies performance‑oriented defaults such as autocast, recommended batch sizes, and CUDA/cuDNN flags. This should complement and build on top of the work tracked in #5 (Adaptive OS-based support for RTX series).

Why

• Many users will have “standard” RTX setups where a curated set of defaults can provide an immediate speed boost.
• Centralizing these choices in a named profile makes it easier to test, reproduce, and iterate on performance improvements.
• Keeping this as a preset ensures existing conservative defaults remain available.

What to do

• Define a named profile (e.g., rtx-fast) in the existing config system.
• For that profile, specify:
  • Autocast / mixed precision defaults appropriate for RTX.
  • Recommended batch size / gradient accumulation settings.
  • Any CUDA/cuDNN knobs or environment variables that are safe and beneficial.
• Expose this profile via CLI/API (for example, --profile rtx-fast), reusing the central config object from Centralize config via Pydantic/dataclass and env overrides #2.
• Add a short section in the README explaining when to use this preset and its trade-offs.

Acceptance criteria

• Users with common RTX GPUs can enable the fast preset with a single flag or config option.
• Measurable speed improvements over the default profile on typical RTX hardware, with no regressions in correctness.
• This issue is linked conceptually (and in description) to Adaptive OS-based support for RTX series #5 so that future GPU work has a clear place to plug in.