ROCm is just a "CUDA" alternative, isn't it?

First of all, **congratulations** for the *ROCm* merge/availability on main, that's a great news/feature for everyone on Strix Halo ... and yet, this is the "future me" asking about some consistency around that target, specially because it just landed so that changing sooner than later will impact way less users.

### In a nutshell

* **CUDA** is handled as high-level target:
* `make cuda-generic` targets RTX based platforms
* `make cuda-spark` targets GDX Spark
* **CPU** is also handled as high-level target, the outcome depends on the current *CPU* capabilities/extensions, I suppose, it's all good
* **ROCm** is handled in a completely different way ... but does it need to?
* `make rocm-generic` *should* target any Radeon GPU capable enough to deal with LLM
* `make rocm-halo` (or *rocm-strix-halo*) *should* target the *gfx1151* brand/branch? that works for *8065S* too, right?

and then again, thanks a lot for this latest merge, I am just thinking if the goal in here is to target those main targets, let's try to be explicit and aligned around the `make` meaning, if we can? Nothing a good README couldn't solve, just wondering why `make` couldn't be more consistent with goals/intents of each command 👋