[Feature Request] Memory map .ort files by default

Describe the feature request

.ort files - being flat buffer-based - are more amenable to being memory mapped and consumed in-place. But - by default - the OnnxRuntime loads an .ort file by reading it into memory. This forces the entire file to be accessed up-front, and loads the file into private memory pages. Memory mapping the file would defer accesses and allow for the opportunity of memory pages to be shared should the model be used by multiple sessions or processes.

Describe scenario use case

Using OnnxRuntime to run inferencing using large models, allowing for the opportunity of lazily loading the model and sharing memory pages of constant tensor data.

No source files, tests, or entry points are named. Start by locating the default .ort model-loading path and its existing loading tests, then determine how memory mapping would affect lazy access and sharing. Done means .ort models are memory mapped by default without breaking existing inference behavior, with coverage for the new loading path.