refactor: remove H5 storage backend in favor of mmap bin

- Remove H5Store, H5Writer, save_h5/load_h5 and h5py dependency
- MmapStore (bin) is the sole pre-tokenized storage backend
- Move setup_logging after imports to fix E402 in __init__.py
- Clean up unused imports across test files
- Move inline test imports to file top

scripts/tools/generate.py

@@ -145,9 +145,7 @@ def processor(
 @click.option("--top_p", type=float, default=0.95, help="Top-p filtering.")
 @click.option("--batch_size", type=int, default=1, help="Batch size.")
 @click.option("--num_samples", type=int, default=1, help="Responses per prompt.")
-@click.option(
-    "--max_seq_len", type=int, default=2048, help="KV cache length."
-)
+@click.option("--max_seq_len", type=int, default=2048, help="KV cache length.")
 @click.option("--frequency_penalty", type=float, default=0.0, help="Frequency penalty.")
 @click.option(
     "--rep_window", type=int, default=64, help="Window size for frequency penalty."

scripts/tools/preprocess.py

@@ -1,4 +1,4 @@
-"""CLI: JSONL → tokenized .h5/.bin via config-driven Pipeline."""
+"""CLI: JSONL → tokenized .bin via config-driven Pipeline."""
 
 import click
 
@@ -8,7 +8,7 @@ from astrai.preprocessing.pipeline import Pipeline
 
 
 @click.command(
-    name="preprocess", help="Tokenize and pack raw JSONL data into .bin/.h5 format."
+    name="preprocess", help="Tokenize and pack raw JSONL data into .bin format."
 )
 @click.argument("inputs", nargs=-1, type=click.Path(exists=True), required=True)
 @click.option(
@@ -30,7 +30,7 @@ from astrai.preprocessing.pipeline import Pipeline
 )
 @click.option("--batch_size", type=int, default=None, help="Records per batch.")
 def preprocess_command(inputs, output_dir, pipeline_config, tokenizer_path, batch_size):
-    """Tokenize and pack raw JSONL data into .bin/.h5 format."""
+    """Tokenize and pack raw JSONL data into .bin format."""
     config = PipelineConfig.from_file(pipeline_config)
     if batch_size is not None:
         if batch_size < 1: