[CI Exp]: CI Revamp

[abukhoy]

2-Phase CI Pipeline

tiny_model: Per PR CI : Expected Time 30 - 40 mins.

Created mainly 4 markers such as:

1. LLM : Tests which are mainly non-vlm models such as causal, embedding, audio, sequence models and its feature run on QAIC.
2. VLM : Tests which are related to image text to text models which run on QAIC.
3. Difussion : Marked which related to diffusion models which run also on QAIC.
4. Non QAIC : All those tests that run w/o QAIC.

Phase 1 (parallel): 4 non-QAIC stages run concurrently

• Non-QAIC: LLM (export/compile)
• Non-QAIC: VLM (export/compile)
• Non-QAIC: Diffusion (export/compile)
• Non-QAIC: Other

Phase 2 (sequential): 3 QAIC stages run one after another for generation

• QAIC: LLM (cards 0-3)
• QAIC: VLM (cards 0-3)
• QAIC: Diffusion (cards 0-3)

QEFF_HOME is shared between Phase 1 and Phase 2 per model family so that ONNX exports produced in Phase 1 are reused by the QAIC compile/generate steps in Phase 2 without re-exporting.

One stage failure never blocks other stages (catchError isolation). TEST_PROFILE parameter controls which model set is used (tiny_model / full_model).

full_model : Nightly CI

[abukhoy]

following up in #1180