[CPU] Expose more options to CPUCodegenOptions. (#23586)

The revision moves `useFastMinMaxOps`, `skipIntermediateRoundings`,
`useSoftmaxInterFusion`, and `instrumentMemoryAccesses` to
`CPUCodegenOptions`, and hides the remaining developer CLI flags.

There are three categories of use cases:

- **Users**: They do not need to understand individual flags. The
expectation is that they use optimization levels and optional flags
recommended by core developers.
- **Core developers**: Internal knobs for debugging or testing specific
codegen behaviors. These are hidden from `--help` and may change or
disappear at any time.
- **External developers**: Contributors pushing the boundary on specific
targets (e.g., ARM SVE/SME). Their flags need discoverability but are
not yet stable.

Moving flags related to scalable vectors was considered but ultimately
not done because the plumbing would violate layering. For example,
`clEnableScalableVectorization` is accessed globally through a free
function. It cannot be passed to the encoding materialization pass
because that pass should not carry target-specific codegen options. The
root cause is that this information is not queryable from the IR.
Promoting it to `CPUCodegenOptions` with an `experimental-` prefix was
considered, but the layering violation makes it the wrong approach.
Thus, it is left where it is.

Signed-off-by: hanhanW <hanhan0912@gmail.com>