fix: generate correct top_k columns when no shared expert

[Mikezhang001]

Bug

When running jit_moe.py with --no-shared-expert, topk_ids is
generated with top_k-1 columns instead of the expected top_k columns.

This happens because generate_topk_ids is always called with top_k-1,
but the if args.shared_expert block that appends the missing column is
skipped when there is no shared expert.

Fix

Move the generate_topk_ids call inside the if/else block:

• With shared expert: generate top_k-1 routed experts, then append
  1 fixed shared expert (index E-1) → total top_k columns
• Without shared expert: generate full top_k experts from all E
  experts → total top_k columns

[Mikezhang001]

HI @SzymonOzog, I noticed a bug in jit_moe.py when using --no-shared-expert.
Could you take a look at this fix when you have time? Thanks!