Expose n_epochs on novel_train and simple_mlp_train

[rcannood]

Depends on #56 -- branched off it, so please merge that one first.

A long-running method should expose whatever bounds its work and cap it for CI,
rather than hardcoding it. novel had for i in range(100) buried in
helper_functions.py, and simple_mlp took its epoch count from the bundled model
yaml, so neither could be sped up for a test run or tuned by a user.

• novel_train: --n_epochs (default 100) bounds the training loop
• simple_mlp_train: --n_epochs overrides the value in the bundled model config
  when given, otherwise the yaml still wins
• Both cap at 2 during viash test via info.test_default, so the real defaults
  are untouched

Checked what the harness actually resolves rather than inferring it from the runtime:

scenario 'run': --n_epochs resolves to 2

before  novel_train        EXIT=0 WALL=37s
after   novel_train        EXIT=0 WALL=13s
before  simple_mlp_train   EXIT=0 WALL=12s
after   simple_mlp_train   EXIT=0 WALL=12s

(simple_mlp was already only 10 epochs, so its test time is all image build.)

Also in scripts/create_datasets/test_resources.sh:

• Each step is skipped when its output exists and is no older than the dataset's
  state.yaml, so re-running after a partial failure doesn't redo everything.
  FORCE=1 regenerates.
• The model fixtures train for 2 epochs -- they only need to exercise the predict
  components, not be any good.

The up_to_date helper is checked against the cases that matter:

output missing                        REBUILD
output newer than inputs              SKIP
one input newer than output           REBUILD
empty directory output                REBUILD
non-empty dir newer than inputs       SKIP
non-empty dir older than an input     REBUILD
FORCE=1 with fresh output             REBUILD