feat: pwm.grad and pwm.grad.ism per-bp gradient vtracks - #111

Implements the linearized MAX-mode gradient (Task 3 of the pwm.grad plan)
for the simplest case: bidirect=FALSE, strand=1, no spatial weighting.

- Precompute m_worst_col0_fwd / m_worst_col_last_fwd at construction.
- Route GRAD_MAX through a dedicated score_grad_max() that scans the
  iterator-clamped anchor range [i_min, i_max] (matching pwm.max's
  argmax convention) and returns M[0, b_p] - worst_col0 at the head
  anchor, 0 elsewhere.
- Reject bidirect=TRUE / strand=-1 with verror BEFORE the try/catch in
  score_interval so the error reaches R instead of being squashed to NaN.
- Skip the sliding-window dispatch for all four gradient modes; defer
  spatial weighting to Task 9 (NaN for now under m_use_spat).

Tests verify both head=argmax (positive gradient matching the oracle) and
head!=argmax (engine returns 0, oracle agrees), plus the two error paths.

Unifies the GRAD_MAX and GRAD_LSE answer paths under score_grad_linearized,
using score_forward_original / score_reverse_original so the rc'd target
(when strand_mode == -1) is canonicalized to fwd-genome semantics. Bidirect
combines per-strand head scores via softmax. Tie-break in argmax favors the
head anchor to match the oracle's R which.max convention.

Extends score_grad to handle ism=true: at the head anchor, only the head's
per-anchor score changes under a base flip, so 3 alternative aggregates can
be computed in O(1) using the pre-scanned best_no_head and lse_no_head.
Covers both LSE and MAX aggregations and all strand modes (bidirect, fwd,
rc). Renames score_grad_linearized to score_grad and unifies the four grad
modes through one function.

Tests: ISM-vs-linearized divergence (argmax shift on a periodic test seq),
ISM bidirect on asymmetric PSSMs, ISM strand=-1, both LSE and MAX.

Per-anchor spat_log is added to each anchor's score in the scan; ISM uses
the head's spat_log to keep the flipped per-anchor score on the same scale
as best_no_head and lse_no_head. Tests: spat_factor=1 invariance and
non-trivial spat with both lin and ism.

… cases

New vignettes/PWM-Functions.Rmd covering the existing PWM aggregating
tracks (pwm, pwm.max, pwm.max.pos, pwm.count), edit-distance variants,
and the new per-position attribution tracks pwm.grad / pwm.grad.ism.
Math sections evaluate as pure-R demos (no genome DB needed); genome
examples use eval=FALSE.

Adds an articles: section to _pkgdown.yml listing the new vignette
alongside the existing ones.

Cache per-anchor (fwd, rc) raw scores in a deque. On consecutive iterator
steps with matching chrom/strand/window, slide the deque (pop trailing,
score+push leading) instead of doing a full O(W*L) rescan. The aggregate is
still recomputed from the deque per pivot (O(W)), so ISM_MAX which needs
best_no_head doesn't require a second monotonic deque.

Bench (HOMER.CTCF, mm10 chr1:3M, 100 kb, W=500, iter=1):
  pwm        0.11s  (1.00x)
  pwm.grad   1.47s (13.25x; was 66.58x)

Full O(1) running aggregates (RunningLogSumExp + dual RunningMaxDeque) would
buy another ~5-10x but require a sliding-window 'max excluding the head' deque
that's tricky for ISM_MAX.

- Add a dedicated 'Strands and bidirect' section that lays out the
  per-anchor score formulas, the bidirect=TRUE strand-union, and the
  position-reporting convention (pwm.max.pos sign, grad pivot at
  interval.start in fwd-genome coords).
- Drop the 'Why LSE rather than just MAX?' subsection.
- Trim gradient-directed framing in the intro and pwm/pwm.max sections.
- Shorten the gradient section's motivation; one-line strand handling
  reference points to the new section.

The PWM is a linear log-likelihood model, so the 'gradient' is just a
PSSM column lookup; integrated-gradients / DeepLIFT contribution rules
don't apply (they collapse to gradient × input for linear models).
ISM is standard in motif analysis (e.g. motifbreakR for SNPs).

Renaming the user-facing description to 'per-bp PSSM-column
contribution' (with explicit 'softmax-weighted' for LSE and
'argmax-conditioned' for MAX) is more accurate than 'gradient /
saliency / DeepLIFT-style attribution'. The tracks themselves stay
named pwm.grad / pwm.grad.ism.

PWM_GRAD / PWM_GRAD_ISM are anchored at seq_interval.start, so the
multi-part filter aggregation that summed scores across disconnected
unmasked parts mixed gradients at different genomic anchors. The single-
part fallback had the same problem when the mask covered the pivot.

Score only the unmasked part that starts at seq_interval.start; return
NaN when the pivot is masked. Add filter tests for both pwm.grad and
pwm.grad.ism (mask pivot, multi-part mask, non-intersecting mask).