[llvm] [RISCV] Introduce local peephole to reduce VLs based on demanded VL (PR #104689)

[Camel Coder via llvm-commits]

camel-cdr wrote:

AFAIK [ara](https://github.com/pulp-platform/ara) and [saturn](https://github.com/ucb-bar/saturn-vectors) are both able to dispatch based on VL.
I don't have the benchmarks for ara, but for saturn, you can see that the rvv_128 variants, perform the same reguadless of LMUL: https://camel-cdr.github.io/rvv-bench-results/saturn/utf8_count.html

That beeing said, both of those are in-order cores that do a bunch of chaining, I don't expect more high performance ooo cores to implement this.

https://github.com/llvm/llvm-project/pull/104689