fix: make Forecaster runtime device-safe

[gyx09212214-prog]

Summary

• choose float16 only when CUDA is available and default to float32 on CPU-only runtimes
• make GPU memory reporting tolerant of missing CUDA/NVML instead of failing the app import path
• add FINGPT_TORCH_DTYPE override support with explicit validation, plus unit coverage for dtype resolution, NVML handling, and CUDA cache cleanup

Why this matters

The Forecaster demo currently assumes CUDA fp16/NVML availability. That can break local CPU-only or Windows smoke tests before users reach the actual forecasting flow. Keeping runtime hardware checks best-effort, while failing fast on invalid explicit dtype overrides, makes the demo easier to validate without changing the model path for CUDA users.

Scope

This addresses the NVML and CPU fp16 runtime failure patterns reported in #119 without changing model loading, prompts, market data access, or generated forecasting behavior.

Validation

• python -m pytest fingpt\FinGPT_Forecaster\tests\test_runtime_utils.py -q - 8 passed
• git diff --check - clean
• Full Gradio/model inference was not run because it requires Hugging Face credentials, the Llama-2 base model, LoRA weights, and market data API keys.