Move pipeline parameter logic into AutoMLAlgorithm

Related to [this discussion](https://github.com/alteryx/evalml/pull/3373#discussion_r832473688).

With [this PR](https://github.com/alteryx/evalml/pull/3373) we have consolidated most of the pipeline parameter-creating logic into the `AutoMLAlgorithm` class. However, there are still cases, like with the sampler parameters, that might be useful to move out of `AutoMLSearch`. My main concern is if we choose to move this, we would need to pass all of the relevant data to `AutoMLAlgo`, like `sampler_name`, `problem_type`, and the `X, y` data.

This issue tracks determining if there is another way to move the pipeline parameter definitions to `AutoMLAlgorithm` in a clean way, and implementing it if so. If not, it would be nice to write about what the limitations/tradeoffs are.

fyi @freddyaboulton @jeremyliweishih