ReDrafter support for Qwen

[darraghdog]

Support the ReDrafter functionalities used in AIMO2,

• With fp8 base model.
• With Qwen base model, and add qwen-7b example in README.
• Break convert_checkpoint.py into two steps. The idea is to give more flexibility around the base model used.
  1/ convert or quantise base model with normal convert_model.py or quantise.py specific to the base model.
  2/ convert redrafter and attach to converted base model
• [SOLVED] Issue using bfloat16 in redrafter build - if we get a base model with a bfloat16 layer, covert to float16, this is switched to float16.
• Move main ReDrafter logic to a mixin
• Add new classes, ReDrafterForLLaMALM, ReDrafterForQWenLM... more can be added as needed, just need to test each one - I may add some more later, eg. Deepseek.
• Updated redrafter pytests accordingly for the above.

Tests done on README examples, attaching logs from tests here :
redrafter_pr_test_logs.txt
.

Update on tests draft_len 6, beam_width 6:
redrafter_pr_beam6_draftlen_6.txt

[FIXED] Update2 on tests draft_len 6, beam_width 6:
redrafter_pr_beam6_base_beam1.txt

[darraghdog]

/bot run

[darraghdog]

/bot run

[rakib-hasan]

/bot run

[tensorrt-cicd]

PR_Github #8552 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #8552 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6203 completed with status: 'FAILURE'

[darraghdog]

/bot run

[rakib-hasan]

/bot run

[tensorrt-cicd]

PR_Github #9897 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9897 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #7304 completed with status: 'FAILURE'

[darraghdog]

@rakib-hasan the build_redrafter_engines.py should be fixed now.
Would it be possible to rebase the branch and run CI ?

[rakib-hasan]

/bot run

[tensorrt-cicd]

PR_Github #10170 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #10170 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #7508 completed with status: 'SUCCESS'