[type] [opt] Use BitStructStoreStmt for CustomFloatType with non-shared exponents

[xumingkuan]

Related issue = #1905 #2128

This PR adds support of CustomFloatType with non-shared exponents in BitStructStoreStmt.

Now the codegen workflow of BitStructStoreStmt looks like the follows:

1. Generate the code for all shared-exponent floating type stores (do not generate the code for non-shared exponent floating type stores here).
2. For each other store, generate the code, and bit_or it to bit_struct_val. For CustomFloatType with non-shared exponents, we need to generate the code of storing both the exponent and the fraction (code mainly copied from visit(GlobalStoreStmt *)).
3. Generate the mask of the bits stored in step 2.
4. If the mask is full, or if the BitStructStoreStmt itself is not atomic (probably because it's the only access to a loop index), generate a masked store; otherwise, generate an __atomic_compare_exchange.

We can remove the support of custom float type in GlobalStoreStmt to make the codebase cleaner in the future in this PR.

[Click here for the format server]

---

[yuanming-hu]

Awesome! LGTM. Thank you very much :-)