chore: trim examples to qwen3 only

[aoshen02]

This PR removes the public example set down to Qwen3-4B and Qwen3-30B-A3B, plus the related OPD and low-precision example/docs entries we already agreed to drop.

It is intentionally separate from the slime->vime rename PR so the rename can be rebased cleanly after this lands.

[gemini-code-assist]

Code Review

This pull request performs a major cleanup by removing outdated documentation, example scripts, and associated tests for features like low-precision training, on-policy distillation, multi-task evaluation, and several specific model configurations. It also introduces minor backend fixes, such as disabling temperature scaling for value outputs in Megatron loss calculations, and adds a new unit test test_loss.py. The review feedback correctly identifies a potential test flakiness issue in the new unit test where sys.modules.setdefault is used to register mock modules, and provides a solid suggestion to directly assign to sys.modules instead.

[gemini-code-assist]

Using sys.modules.setdefault to register mock modules can lead to flaky or failing tests if another test in the suite has already imported megatron (since setdefault will not overwrite the existing entry in sys.modules). Directly assigning to sys.modules ensures that this unit test always uses the intended mock modules regardless of the test execution order.

Suggested change

	sys.modules.setdefault("megatron", megatron_mod)
	sys.modules.setdefault("megatron.core", megatron_core)
	sys.modules["megatron"] = megatron_mod
	sys.modules["megatron.core"] = megatron_core

[CalvinXKY]

LGYM. Ensure it doesn't affect the CI.