avoid redundant memory allocation for external initializers

[frank-dong-ms]

Description

avoid redundant memory allocation for external initializers, we will use mmap for external initializers later so no point to allocate memory in advance then release them later.

Motivation and Context

In current implementation, we will:

1. Allocate memory (with desired size of current initializer) for initializer first: https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/framework/session_state_utils.cc#L131
2. For external initializer, we will point initializer to mmaped object in memory and release previously allocated tensor: https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/framework/session_state_utils.cc#L89

For large models, we are keep allocating and release memory for external initializers which seems unnecessary.

For phi silica model, with this change we can reduce transient memory usage from 4,566MB to 2,724MB. Since these redundant memory is released quickly when we mmap external initializers so this change has no much impact on peak memory usage.