Reject sub-MIL-floor cross-compile targets instead of raising

[sbryngelson]

Problem

cross_compile_check(graph, "h11") raised ValueError instead of returning False:

aneforge/_targets.py:278: ValueError: family 1 below the floor for limit 'max_tensor_dim'

preflight() calls limit("max_tensor_dim", family) (and other per-family caps) at the top, but limit() is only defined at or above the H13+ MIL floor (Family.A13=2). For a sub-floor target family — Family.OLDER=1, i.e. h11/h12/m9/t0 — limit() correctly raises, and preflight never reaches its rep.ok rejection path. So cross_compile_check (and compile(target=), which gates on the same preflight) crashed on any below-floor target.

This is the deterministic failure behind tests/test_cross_compile.py::test_relu_rejected_below_the_mil_floor in the on-device corpus.

Fix

Guard preflight() with the existing supports_mil(family) check. Below the floor, report every op as rejected (so rep.ok is False) without querying the floor-only limits, reusing the function's own graph-walk:

if not supports_mil(family):
    # Below the H13+ MIL floor: the e5rt/MIL path won't run at all -> all ops rejected.
    ... walk graph, append each op to rep.reject ...
    return rep

Fixed at the root layer (preflight), so both cross_compile_check and compile(target=) now reject sub-floor targets cleanly.

Note: host capability is unaffected

This only changes the verdict for sub-floor target families being cross-compiled for; it does not gate the host. Verified:

h11 target (family 1): preflight.ok = False   # correctly rejected
M5 host  (family 5):   preflight.ok = True    # fully capable, unchanged

Verification

• The previously-failing test_relu_rejected_below_the_mil_floor now passes.
• pytest tests/test_cross_compile.py tests/test_cross_compile_matrix.py tests/test_compile_targets.py tests/test_targets.py: 141 passed.
• Blast radius is tiny: the new branch only triggers for family < 2, which never occurs for real hardware or normal compiles.

Branched off main, independent of the display/ASCII cleanup in #1.