Honour VisuGroupDepVals valsStart when two frame groups share an array

[gdevenyi]

Fixes #200.

Spec 7.4:

typedef struct {
    char name[VISU_MAX_NAME_LEN];  // Name of dependent parameter (must be an array)
    int valsStart;                  // Start index in the dependent parameter array
} VISU_DEPVALS_TYPE;

frame_group_values read only name and never valsStart. When two groups depend on the same parameter the writer concatenates their blocks and gives each group its own start; recording both axes under one name made the code spread the concatenation over the product of the two axis lengths, which never matched, so the parameter was dropped.

Before / after

resources/testdata/bruker2nifti_qa/raw/Cyceron_DWI/.../3/pdata/2 — FG_DTI (len 22, valsStart 3), FG_CYCLE (len 3, valsStart 0), 25 comments:

before:  VisuCoreDataUnits, VisuFGElemId, VisuCoreOrientation, VisuCorePosition
         (VisuFGElemComment missing -- 25 % (22*3) != 0)

after:   ... plus
         VisuFGElemComment[FG_DTI]    (1,1,22,1,1)  'Fractional Anisotropy', 'Trace', 'Intensity', ...
         VisuFGElemComment[FG_CYCLE]  (1,1,1,1,3)   'R 1', 'R 2', 'R 3'

Change

Each (group, dependency) pair gets its own window parameter[valsStart : valsStart + group_len]. A parameter owned by one group keeps its plain name, so nothing that worked before changes; a shared one is keyed name[GROUP_ID], which is the only way a single mapping can express two windows. The alignment logic moved into _align_to_frame_groups unchanged.

Across 514 corpus reconstructions the result changes for exactly three — Cyceron_DWI/3/pdata/1, 3/pdata/2 and pv5/0.2H2/31/pdata/2 — and every change is an addition; nothing was removed or altered.

Tests

• test_two_frame_groups_can_share_one_dependent_parameter — a synthetic two-group reconstruction whose VisuFGElemComment is shared with different starts. Fails on master with KeyError.
• test_a_parameter_owned_by_one_frame_group_keeps_its_plain_name — guards the unchanged single-group key.

Suite: 2142 passed, 12 skipped.

[gdevenyi]

Rebased onto #213 (ruff format), so this branch carries that commit until #213 merges — merge #213 first and this diff reduces to its own change. Verified after the rebase: ruff check and ruff format --check clean, full suite green.

Integration check across all eleven conformance PRs merged together: 2164 passed, 12 skipped; corpus load test 3202 → 3207 of 3478 with zero newly broken datasets.