Add MDAL_LF_SkipStatistics load flag and approximate min/max API

[nicogodet]

Lets callers opt out of the eager per-dataset statistics scan during MDAL_LoadMesh, dramatically reducing load time for large multi-timestep meshes when an exact min/max is not required upfront.

The gain is substantial!
I tested a "small" SELAFIN result file. We can have bigger ones which freeze QGIS for several minutes...

Fixes #526

Tested on QGIS with improvements to use this new API. Mesh file is stored on USB3 drive stick. I don't have access to network shared drive for now. My hardware is pretty old (i7-4790, 8Gb RAM).

Python snippet

from qgis.core import QgsMeshLayer, QgsProject, QgsSettings
from time import perf_counter

MESH_PATH = "/media/nicogodet/Nouveau nom/v5.res"
SAMPLE_COUNT = 10

s = QgsSettings()
key_on = "mesh/approximate-min-max-on-load"
key_n  = "mesh/approximate-min-max-sample-count"

def load_and_inspect(label):
  t0 = perf_counter()
  layer = QgsMeshLayer(MESH_PATH, f"v5_{label}", "mdal")
  elapsed = perf_counter() - t0
  if not layer.isValid():
      print(f"[{label}] invalid layer")
      return None, elapsed
  QgsProject.instance().addMapLayer(layer)

  provider = layer.dataProvider()
  n_groups = layer.datasetGroupCount()
  print(f"\n[{label}] load: {elapsed:.3f} s   groups: {n_groups}")
  print(f"  {'#':>3}  {'name':<32}  {'kind':<7}  {'ts':>6}  {'min':>14}  {'max':>14}")
  for g in range(n_groups):
      meta = provider.datasetGroupMetadata(g)
      n_ts = provider.datasetCount(g)
      kind = "scalar" if meta.isScalar() else "vector"
      print(f"  {g:>3}  {meta.name()[:32]:<32}  {kind:<7}  {n_ts:>6}  {meta.minimum():>14.6g}  {meta.maximum():>14.6g}")
  return layer, elapsed

# --- exact (toggle off) ---
s.setValue(key_on, False)
s.setValue(key_n, SAMPLE_COUNT)
QgsProject.instance().removeAllMapLayers()
_, t_exact = load_and_inspect("exact")

# --- approx (toggle on) ---
s.setValue(key_on, True)
QgsProject.instance().removeAllMapLayers()
_, t_approx = load_and_inspect(f"approx-{SAMPLE_COUNT}")

# --- résumé ---
print(f"\n>>> exact:                 {t_exact:.3f} s")
print(f">>> approx ({SAMPLE_COUNT} timesteps):   {t_approx:.3f} s")
print(f">>> speedup: x{t_exact / t_approx:.1f}  ({t_exact - t_approx:+.3f} s saved)")

Results

[exact] load: 44.877 s   groups: 7
    #  name                              kind         ts             min             max
    0  vitesse       ms                  vector      507               0         5.56107
    1  hauteur d'eau   m                 scalar      507               0          26.833
    2  surface libre   m                 scalar      507           -8.88           68.45
    3  fond            m                 scalar      507           -8.88           68.45
    4  frottement                        scalar      507               6              50
    5  cote maximum    m                 scalar      507           -8.88           68.45
    6  vitesse maximum ms                scalar      507               0         21.0616

[approx-10] load: 2.296 s   groups: 7
    #  name                              kind         ts             min             max
    0  vitesse       ms                  vector      507               0           3.911
    1  hauteur d'eau   m                 scalar      507               0          26.833
    2  surface libre   m                 scalar      507           -8.88           68.45
    3  fond            m                 scalar      507           -8.88           68.45
    4  frottement                        scalar      507               6              50
    5  cote maximum    m                 scalar      507           -8.88           68.45
    6  vitesse maximum ms                scalar      507               0         21.0616

exact:                 44.877 s
approx (10 timesteps):   2.296 s
speedup: x19.5  (+42.581 s saved)

AI disclaimer

Claude code (mostly Opus 5 and reviewed by Fable 5.1) done most of the work. I wrote the specifications (look at what GDAL does for rasters, dig in QGIS code base to learn the load path of rasters and meshes then look at MDAL to implement a similar approximative stats as GDAL does), Claude did the work. I did not reviewed all the changes but I read a bit to correct some comments and reduce verbosity.

[nicogodet]

@uclaros Is it something you want to look at?
If so, please take a look at linked QGIS issue which explain why