[SYSTEMDS-3953] Switch quantile default to R quantile type 7 - #2585

Bring the Spark path in line with the kernel switch to R quantile type 7
in the previous commit. Row-partitioned single-column and weighted
two-column inputs both return type-7 interpolated values matching CP.

QuantilePickSPInstruction:

  * processInstruction reads as a clean switch — VALUEPICK / MEDIAN go
    through pickQuantileValues, IQM goes through the extracted
    computeIqm. The pre-3953 shape mixed the two under a single
    getWeightedQuantileSummary blob whose fields meant different things
    per operation; the split makes each case's data shape explicit.
  * pickQuantileValues handles N >= 1 quantiles uniformly. Weighted
    branch stores rank g's in a primitive double[] and pulls (lo, hi, g)
    triples from extractWeightedTriples; unweighted branch uses
    MatrixBlock.computeType7Rank and interpolates from the collected
    sorted values.
  * computeIqm is the extracted IQM path — reuses extractWeightedTriples
    but combines with computeIQMCorrection, not type 7. IQM behavior is
    numerically unchanged from pre-3953.
  * extractWeightedTriples is the shared helper. Locates the partition
    holding each ceil-based key by scanning cumulative per-partition
    weights, then invokes ExtractWeightedQuantileFunction to fetch the
    (position, posPart, value) triples.
  * Type7Rank inner class and getWeightedQuantileSummary are both
    removed; the static kernel helper subsumes the former, and the
    per-consumer split obsoletes the latter.

Tests: un-ignore testQuartileArrayCP / testQuartileArraySP (marked with
FIXME: fix SYSTEMDS-3953 by David); add QuantileEven{1,2,3}{CP,SP} at
n = 128 for p in {0.25, 0.5, 0.75}. Odd-n cases (rows = 1973) can't
exercise interpolation because the rank lands on an integer at classical
p; the even-n additions cover the g != 0 path in both CP and SP.

FED path in the follow-up commit on this branch.

Bring the FED path in line with the CP/SP switch to R quantile type 7 in
the previous two commits. Row-federated matrices now return the same
type-7 interpolated values as CP: [0.239, 0.517, 0.890, 0.944] at
p = 0.25, 0.5, 0.75 gives 0.4475, 0.7035, 0.9035.

processRowQPick's VALUEPICK / MEDIAN branch computes per-quantile rank
triples (lo, hi, g) via MatrixBlock.computeType7Rank, flattens to a
deduplicated int[] of ranks, feeds those to pickMultipleRanks (renamed
from computeMultipleQuantiles to signal it now does one job: fetch the
value at each rank), and interpolates (1 - g) * v_lo + g * v_hi at the
caller. Decoupling ranks from interpolation lets the multi-rank pipeline
stay type-7-agnostic — an earlier attempt threading an isType7 flag
through createHistogram / getBucketWithIndex regressed VALUEPICK to
type 1 and mis-computed q25 boundaries, so this shape avoids
reintroducing that seam.

Consequence of the type-7 unification: the boolean average thread across
processRowQPick, createHistogram, getBucketWithIndex, and
getSingleQuantileResult (which only ever handled the even-n median
average case) is removed end-to-end.

IQM stays a raw ceil-based trimmed weighted mean and is extracted into
private computeIqm, mirroring the SP computeIqm helper from the previous
commit so processRowQPick reads as VALUEPICK/MEDIAN -> rank pipeline,
IQM -> computeIqm — symmetric with SP.

refineBucket wraps the "recurse into a coarse-histogram bucket with a
finer sub-histogram" step so computeIqm and pickMultipleRanks share the
nextNumBuckets heuristic in one place instead of two.

MEDIAN in processColumnQPick now dispatches through the VALUEPICK path
with p = 0.5 (kernel pickValue(0.5) is already type 7 after the first
commit), which lets the ColMedian inner UDF go away. VALUEPICK and its
ValuePick UDF are otherwise unchanged.

Weighted row-federated inherits the kernel's sum-of-weights extension
for free: N = sumWeights in the two-column case flows through the same
rank formula. Column-federated weighted stays out of scope, matching
PR apache#2497.

Verified: FederatedQuantileTest 24/24, FederatedQuantileWeightsTest
12/12, QuantileTest 32/32 (including newly un-ignored
testQuartileArray{CP,SP} and the QuantileEven cases), IQMTest green,
QuantilePickTest green.