Add paths.buffer config key to redirect buffer snapshot storage

[oleg-savchenko]

Mirrors the existing paths.samples pattern: if paths.buffer is set in the config (or passed as a Hydra override), buffer snapshots are written to {paths.buffer}/snapshots/ instead of the default {run_dir}/buffer/snapshots/.

Useful when run_dir is on a fast local filesystem but buffer snapshots (which can be large) should land on a separate scratch volume.

Motivation

In HPC workflows, adaptive falcon training produces two categories of output with
very different storage requirements:

Output	Typical size	Storage preference
Buffer snapshots (buffer/snapshots/)	Large — grows with max_samples × field_size × adaptive_rounds	Fast scratch (high I/O, temporary)
Posterior samples (samples/posterior/)	Large	Fast scratch
Model checkpoints (graph/)	Small	Persistent home/project dir
Configs, logs, plots	Small	Persistent home/project dir

paths.samples already allows redirecting posterior samples to a separate volume.
However, buffer snapshots were hardcoded to {run_dir}/buffer/snapshots/ with no
override, so they always landed alongside the model and config files — filling up
persistent storage with large temporary simulation data.

This PR adds a paths.buffer config key — mirroring the existing paths.samples
pattern exactly — so buffer snapshots can be independently routed to scratch while
keeping run_dir (model checkpoints, configs, logs, plots) on persistent storage.

Note: paths.buffer was previously present as a dead/unused key and removed in
75ab086. This re-adds it as a working, documented feature.

Change

falcon/cli.py — one new variable, mirrors the paths.samples pattern:

# before
snapshots_path=str(Path(cfg.run_dir) / "buffer" / "snapshots"),

# after
buffer_base = cfg.paths.get("buffer", str(Path(cfg.run_dir) / "buffer"))
snapshots_path=str(Path(buffer_base) / "snapshots"),

[cweniger]

There are merge conflicts, can you please merge with the latest master?