ci: guard coverage combine against empty coverage glob in fast shards

[achartier]

What

Guard the trailing coverage combine .coverage* in all 13 L1_Functional_Tests_*.sh shard scripts so it is skipped when no coverage files were produced.

cd ${PROJECT_ROOT}/tests
if compgen -G ".coverage*" > /dev/null; then
    coverage combine .coverage*
fi

Why

#2345 split the monolithic L1 functional tests into per-domain shards, each ending with an unconditional coverage combine .coverage*. Under CI:Lfast (FAST=1), run_test skips every sub-test not marked fast. The AutoModel and Megatron_3 shards have zero run_test fast tests, so nothing runs and no .coverage* files are produced. The glob then expands to the literal .coverage*, coverage combine exits non-zero ("Couldn't combine from non-existent path"), and with set -euo pipefail the whole shard aborts → the harness reports "Did not finish." → FAILURE.

This makes fast_L1_Functional_Tests_AutoModel and fast_L1_Functional_Tests_Megatron_3 red on every CI:Lfast PR since #2345 merged. The other shards survive only because they each happen to have ≥1 fast test producing a coverage file.

Fix scope

The broken-today shards are AutoModel and Megatron_3, but the identical unguarded tail exists in all 13 shards, so the guard is applied uniformly to prevent the same latent failure if the last fast test is ever removed from another shard.

The guard still surfaces genuine coverage combine failures when coverage files do exist — it only no-ops the empty case.

Labeled CI:Lfast so the previously-failing AutoModel/Megatron_3 fast shards exercise this fix directly.

cc @chtruong814 (author of #2345)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[achartier]

/ok to test a45a19b

[chtruong814]

@achartier thanks for this! I was working on the fix as well and saw your PR.