[CI] Add multi-arch CI workflows for Windows

[ScottTodd]

Motivation

Fixes #3325. Now the multi_arch_ci.yml workflow builds for Windows too.

Technical Details

Notable differences from build_windows_artifacts.yml:

• No build cache (ccache or otherwise). We really need [CI] Use windows cluster bazelremote service for builds #2415 for that. We could use github actions cache with a unique key scheme per build stage but I'd rather not do all that extra plumbing when we know the better long term solution
• No "post build upload" of logs or index pages (same as multi-arch linux workflows)

Test Plan

Test run triggered: https://github.com/ROCm/TheRock/actions/runs/22081356438

Test Result

• First attempt FAILED building the math-libs stage: https://github.com/ROCm/TheRock/actions/runs/22081356438/attempts/1 (cancelled / timeout / runner disconnect?)
• Second attempt succeeded: https://github.com/ROCm/TheRock/actions/runs/22081356438/job/63901076696

Build timing (approximate, sample size of 1):

Job	Total	Pre-build	Build	Post-build	Overhead %
Linux
foundation	4m57s	1m39s	3m9s	8s	36%
media	1m52s	1m25s	22s	4s	81%
compiler-runtime	18m54s	3m4s	15m36s	14s	18%
profiler-apps	7m23s	1m15s	6m3s	5s	18%
dctools-core	4m47s	1m26s	3m15s	5s	32%
comm-libs (gfx94X)	16m20s	1m13s	15m2s	5s	8%
math-libs (gfx94X)	95m33s	4m26s	90m47s	20s	5%
Windows
foundation	7m5s	5m45s	1m2s	17s	85%
compiler-runtime	26m17s	7m39s	17m55s	42s	32%
math-libs (gfx110X)	206m2s	10m43s	194m44s	35s	5%

Overhead is quite high for the foundation stage. Moving install steps into the base VM image will be the biggest help.

Build performance is poor as expected, at least until ccache is wired up.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.