[CI] Refactor CI for better caching

[brnorris03]

Problem: Unnecessary rebuilding of the same tt-mlir; Docker container builds time out as a result but also (some) regular CI jobs take longer than they should. Inconsistent caching.

Solution:

• Add call-build-ttmlir-toolchain.yml as dedicated cache builder
• Simplify CI workflow to restore cache with fail-on-cache-miss
• Simplify container workflow to use same unified cache
• Update Docker build to require pre-built toolchain via --ttmlir-toolchain
• Update Dockerfile to use toolchain from build context
• Make ccache caching actually work by caching it per-ref not per commit
• Add CI_WORKFLOWS.md documentation, update BUILD_SYSTEM.md, containers/README.md with new architecture
• Refactor common functionality into helper scripts

New on-pr workflow:

  on-pr.yml (triggers on PR)                                                                      
      │                                                                                           
      ├── pre-commit (parallel)                                                                   
      │                                                                                           
      ├── toolchain (check-cache → build if needed)                                               
      │       │                                                                                   
      │       └── build (waits for toolchain via needs:)                                          
      │                                                                                           
      └── test-sim (parallel)                                                                     
      │                                                                                           
      └── check-all-green (waits for all) 

Follows GitHub Actions best practice: dedicated workflow builds and caches dependencies, other workflows consume with fail-on-cache-miss.

When the toolchain cache needs rebuilding, the regular on-pr workflow fails with a clear error message and instructions, e.g.:

Error: Toolchain cache not found for tt-mlir commit dd15572dc0b92167ea1b186161ea4f74107b0329
============================================================
  TOOLCHAIN CACHE NOT FOUND
============================================================
The LLVM + tt-mlir toolchain cache does not exist for this commit.
To fix this, run the toolchain build workflow:
  1. Go to: https://github.com/tenstorrent/tt-lang/actions/workflows/call-build-ttmlir-toolchain.yml
  2. Click 'Run workflow' button
  3. Click the green 'Run workflow' button in the dropdown
  4. Wait for completion (~3-4 hours for full LLVM build)
  5. Re-run this workflow
Cache details:
  Key: Linux-ttlang-toolchain-v1-dd15572dc0b92167ea1b186161ea4f74107b0329
  tt-mlir commit: dd15572dc0b92167ea1b186161ea4f74107b0329
See .github/CI_WORKFLOWS.md for more information.
============================================================

Checklist:

• Self-reviewed (style, logic)
• Added tests (or justified none needed)
• PR is small and focused (one task)