fix: use OrderedDict for proper LRU cache eviction in fp8_cast_bf16.py

[modimihir07]

Problem

The loaded_files cache in fp8_cast_bf16.py uses a plain Python dict with
FIFO eviction (next(iter(loaded_files))). The comment on line 90 says
"keep only the 2 most recently used files", but the implementation is actually
FIFO (first-in-first-out), not LRU (least-recently-used).

When get_tensor() accesses a cached file (e.g., to fetch a scale_inv tensor
that lives in a different shard than its weight), that access does NOT update
the file's position in the eviction order. This means:

1. File A is loaded (contains weight)
2. File B is loaded (contains scale_inv for a weight in File A)
3. File C is loaded → eviction removes File A (FIFO order)
4. Another weight needs its scale from File A → reload File A, evicting File B
5. This cascades into O(n²) redundant file reloads

Fix

• Replace dict with collections.OrderedDict
• Call move_to_end() on cache hits in get_tensor() to mark as recently used
• Call move_to_end() when loading current file in main loop
• Use popitem(last=False) for proper LRU eviction
• Change if len > 2 to while len > 2 to handle edge cases

Testing

• python -m py_compile inference/fp8_cast_bf16.py — passes
• Change is backward compatible (no API changes)
• OrderedDict is from Python stdlib, no new dependencies