Support microscaling (mx) type

[jinevening]

What

Let's introduce mx (microscaling) type to circle.

Since we don't have any backend that uses mx type yet, let's focus on adding new types to circle schema rather than implementing mx type kernels.

This issue will be closed after the following goals are achieved.

1. Mx type circle can be created from recipe
2. MX type circle can be imported as luci graph
3. luci graph with mx type can be exported as circle
4. MX type circle can be visualized by circledump.

Deadline: 11/30

[jinevening]

#14294 addresses the issues to achieve the goals in the main comment, but there are some remaining issues.

1. MX type axis

• Even if tensors have the same MX type, their axes can be different (the axis would be determined by consumer operator). For example, axes of BMM Op's inputs are reduction dimension. So, for A @ B, A's axis would be -1, and B's axis would be -2.
• We may need a way to describe the axis, e.g., MXQuantizationParameter.

2. luci-interpreter

• If we'd like to do value check in luci-interpreter, we may need to extend luci-interpreter::Tensor class (and memory manager that is in charge of alloc/dealloc of Tensor) because it does not have data structure to save shared exponent.

3. Circle type inference

• Circle's type inference rule for BMM enforces input dtype == output dtype. This constraint may become a problem when we should support mixed-precision operators.

[jinevening]

The goals described in the main comment are all achieved. Let's close this issue.

Further issues listed in #14293 (comment) (and there may be more) will be addressed when necessary.