Revisit generated type for binary bytes

If we get a binary bytes,

in RLC layer,
if it's in input model, we should generate it as `string | Uint8Array | ReadableStream | NodeJS.ReadableStream`,
if it's in output model, we will generate it as `Uint8Array`

in Modular layer,
current implemtation is to treat it as `Uint8Array` regardless input or output.

This could be problematic if there's a large file to upload in the Modular layer. And I am a little concerned if Uint8Array is appropriate if there's a large file to download in RLC ?

we need to revisit how to handle binary bytes in both RLC and Modular in the context of input and output.

Also, do we need to revisit the decision on whether splitting models as input and output in the modular layer ? Given there's a similar discussion about `unknown` being handled differently in the RLC layer between input and output and we are struggling about how to express `unknown` in the Modular layer. see details from https://github.com/Azure/autorest.typescript/issues/2078

/cc @xirzec @joheredi