Add int8 shared-block bigram+cache prototype

[sauravtom]

Summary

• add shared-block transformer support (12 logical layers from a 2-block bank) plus LSQ-lite QAT hooks in train_gpt.py
• add bigram + short cache logit fusion, label smoothing, EMA/SWA options, and int8 export extras for priors/config
• add submission folder records/track_10min_16mb/2026-03-25_Int8_Bigram_Cache_Proto with README, submission.json, train.log placeholder, and runnable train script copy

Testing

• python3 -m py_compile train_gpt.py

Notes

• This PR is a prototype submission package while compute-backed full 8xH100 run logs are pending.