fix: use chunked reading for SHA256 to prevent OOM crash on large model files

[amathxbt]

Bug

check_model_files() in demo/nesa/download.py reads entire model files into memory before computing their SHA256 checksum:

# Before (broken)
with open(output_folder / sha256[i][0], "rb") as f:
    bytes = f.read()  # loads entire file into RAM
    file_hash = hashlib.sha256(bytes).hexdigest()

Impact: Model files are routinely 4–70 GB. Loading one fully into RAM before hashing causes an out-of-memory crash on most consumer hardware, making integrity verification completely unusable.

Also note: the variable is named bytes, shadowing the Python built-in.

Fix

Replaced with hashlib's incremental update API using 8 KB chunks — identical hash output, O(1) memory usage:

# After (fixed)
with open(output_folder / sha256[i][0], "rb") as f:
    hasher = hashlib.sha256()
    for chunk in iter(lambda: f.read(8192), b""):
        hasher.update(chunk)
    file_hash = hasher.hexdigest()