performance optimized

[quangtuyen88]

1. Fixed Synchronous File I/O in modal-login/app/db.ts

Changes:

• Converted all file operations from fs.readFileSync/fs.writeFileSync to async fs.promises
• Made all exported functions async: upsertUser, getUser, getLatestApiKey, setApiKeyActivated
• Added parallel reads/writes with Promise.all() in upsertUser (reads both files concurrently, writes both concurrently)
• Proper error handling for ENOENT (file not found) cases

Impact: Eliminates event loop blocking, allows concurrent request handling, dramatically improves server throughput under load.

2. Batched Model Inference in rgym_exp/src/trainer.py

Changes:

• Rewrote _get_choice_logits to process all choices in a single batched forward pass
• Tokenizes all choices upfront and pads them to the same length
• Stacks sequences into a batch tensor
• Single model forward pass instead of N separate passes
• Manually computes per-sample log probabilities from batch output

Impact: Reduces from N model forward passes to 1 (where N = number of choices). For 4 choices, this is a ~75% reduction in inference time. For 10 choices, ~90% reduction.

3. Optimized Nested Loops in web/api/dht_pub.py

Changes:

• Pre-cache peer names for all peers at once (single batch lookup)
• Reuse single timestamp for all messages instead of calling datetime.now() repeatedly
• Extract peer name from cache using O(1) dictionary lookup instead of repeated function calls
• Cleaner hash key construction (same MD5 operation but more explicit)

[quangtuyen88]

@jcd496 can you help review this pr ? i have tested and all is working fine.