Configure FL backends before TP overlap initialization

[lihongyang1990]

Problem

When tp_comm_overlap is enabled, initialize_megatron() creates TransformerEngine UserBuffers before returning to pretrain(). The existing TE_FL_* and MG_FL_PREFER environment variables were only set afterward, so calls made by initialize_ub() could resolve and cache a backend before te_fl_prefer took effect.

Changes

• Apply MG-FL and TE-FL backend-selection arguments before distributed setup and UserBuffers initialization.
• Reset the cached TransformerEngine-FL selection policy when it was already loaded, without importing TransformerEngine during configuration.
• Remove the late environment writes from pretrain() while keeping FlagGems enablement in its existing post-initialization position.

With the configuration applied before UserBuffers are created, no post-initialization buffer rebuild is needed. The existing distributed_backend: flagcx flow remains unchanged.