Feature/add modal backend

[alay2shah]

Adds support for running training jobs on Modal's serverless GPUs directly from a Mac or laptop. Add a modal: section to any existing config YAML and run leap-finetune config.yaml — the CLI detects it and submits to Modal instead of running locally.

Key changes:

• Environment markers on Linux-only deps (flash-attn, deepspeed, mpi4py, liger-kernel, turm) so uv sync works on macOS
• Modal added as a regular dependency
• New backends/modal_backend.py handles image building, job submission, and detach mode
• resolve_config_path extracted to its own module so SLURM/Modal paths avoid importing torch
• GPU guard gives a clear error when running locally without CUDA

[alay2shah]

This is complete on my end @Paulescu