Run without diversification field

[PhilippVerpoort]

Allow DREAM/HERMES to run without a diversify field

LEGACY genuinely needs at least one Diversify-mode field, but DREAM and HERMES's swap machinery degrades gracefully to optimising unique meetings only when no diversity fields are configured -- so drop the same upfront guard for those two algorithms.

Doing so surfaces two latent crashes in the now-reachable degenerate path, both fixed here:

• HERMES's pareto_prob_for() called max() on the empty pareto_probs dict whenever a candidate improved no diversity field, which is every candidate once there are no diversity fields at all -- a guaranteed ValueError on the first swap. Falls back to 0.0 instead, which keeps select_key() on its meeting-uniqueness branch, DREAM's own natural behaviour in this case.
• (Investigated, not a real bug: DREAM/HERMES's shared select_key() divides by the summed pareto score of the tied candidates, which looked like a possible zero-division once no diversity fields contribute a nonzero score -- but pareto_swaps()'s own distinct-value collapsing always leaves exactly one candidate whenever every candidate's score is 0, so that branch's division is never reached with >1 candidates. Verified empirically across many seeds/configs.)

Confirmed byte-identical DREAM/HERMES output before/after this patch whenever a diversify field is present -- this only changes behaviour in what used to be exception paths.

Adds regression tests for both algorithms running without a diversify field, and for LEGACY still requiring one.

Fix dtype crash in allocate_pandas() with an empty fields dict

participants[list(fields.keys())].apply(...) on a zero-column selection (fields={}) returns an empty DataFrame with no columns to infer a dtype from, so pandas defaults it to float64. to_numpy() then carries that float64 dtype through to allocate_numpy(), which rejects it with "Argument participants must only contain integers." -- masking whatever the algorithm would otherwise have said (including LEGACY's own, more meaningful "One diversification field required!" for this same input).

Build the (n_participants, 0) array directly, with an explicit integer dtype, when fields is empty, instead of going through to_numpy() on the empty-column selection. Confirmed byte-identical allocate_pandas() output before/after this patch whenever fields is non-empty -- the new branch only fires for the previously-broken fields={} case.

Adds regression tests for DREAM/HERMES allocating with fields={} (needs the DREAM/HERMES-without-diversify-field patch stacked below this one, since fields={} implies no diversify field either), and for LEGACY now raising its own diversification-field error for fields={} instead of the dtype error.