New features

• Support Gemma Model (#1631)
• Support Tensor Parallelism (#1599)

Fixes and improvements

• Avoid initializing unused GPU (#1633)
• Read very large tensor by chunk if the size > max value of int (#1636)
• Update Readme