Feature/calibration data device

[avtc]

@Qubitium Hi, this feature allows specify in config the device where calibration data inputs/outputs will be stored, allowing to use more calibration data samples for quantization, because calibration data can be placed on device different to cuda:0 which already stores all layer modules.

Before the feature initial calibration data was stored on CPU and after first pass it was stored on DEVICE_0 (cuda:0 usually).
After the feature if calibration_data_device is not set initial behavior preserved.
calibration_data_device can be set to "cpu", "cuda:1" (or any other torch device), and to "balanced" - in "balanced" mode calibration data distributed between compute devices available: DEVICE_0 .. DEVICE_N

P.S. I have used this feature previously several times but on another old branch. This PR is based on latest master.
Also I have fixed examples in config file for using moe parameter, and fixed sys.abiflags typo which failed build.

Note: the handling of layer with all modules excluded from quantization was also fixed, as current main code did not do forward replay it seems.

I have run several small tests (few first layers) ensuring nothing fail with auto_forward_data_parallel enabled and disabled, on qwen3-30b-a3b with calibration_data_device set to cpu, cuda:1, balanced and removed from config.

[Qubitium]

@avtc Thanks again for amother gem! Can you whip up some unit tests so there is good test coverage on the diffs so I can run it ok our gpus and check for regressions.

[avtc]

@Qubitium I have added tests with help of GLM-5, please review if it is OK

[Qubitium]

@avtc Will be checking and merging in the next 48 hours.

[Qubitium]

@avtc Sorry this took so long. I am currently doing a refractor which should add some hooks to make this PR less code-heavy. Since GPT-QModel will support gguf amongst other weight-only (no calibration) required quantization methods, I need to heavily refractor the current api so we can handle lots of hooks and less messing with lifecycles.

[Qubitium]

@avtc Pleae add/invite me to your PR branch so I can push changes. I will need to merge with v6.0/main once it's ready and do refractoring.

[avtc]

@Qubitium I have 'Allow edits from maintainers' enabled, so you should have write access to this branch. Feel free to push the refactoring and merge with v6.0/main whenever you're ready!

[Qubitium]

Status: refractor/merge pending this v6.0 branch stable:

#2456