fix(api): validate and normalize request seeds

[sufubao]

Summary

• expose omitted or null request seeds as None in the OpenAI-compatible request models
• normalize None to the internal -1 random-seed sentinel
• accept only -1 or non-negative seeds through 2^63 - 1
• validate before assigning to the signed ctypes.c_int64 field, preventing oversized Python integers from silently wrapping
• keep seed normalization and validation helpers alongside SamplingParams

PyTorch accepts seeds through 2^64 - 1, but LightLLM stores request seeds in a signed c_int64; this PR uses the largest value that can be represented without conversion or wraparound.

Validation

• pre-commit (black and flake8) on all modified production files
• manual boundary smoke check for None, -1, 0, 2^63 - 1, underflow, overflow, and very large Python integers