Add correlated K/Kpri scoring and scatter-dominant mode to egs_kerma

[mainegra]

Summary

• Adds correlated $K/K_\text{pri}$ output with statistically correct uncertainty,
  via a per-region cross-term array scg_kp[j][ir] $= \sum_i K_i \cdot q_i$
  accumulated once per history in endHistory.
• Adds a second scoring mode (scatter correction = no) that scores
  $K_\text{scat}$ (kerma from latch $\neq$ 0 photons) instead of $K_\text{pri}$,
  and derives $K/K_\text{pri} = K/(K - K_\text{scat})$.
  Default (scatter correction = yes) scores $K_\text{pri}$ (latch $=$ 0
  photons) directly.

Physics

Both modes report $R = K / K_\text{pri}$. They differ in what secondary
quantity $q$ is scored and how $K_\text{pri}$ is obtained from it.

Scatter-correction mode (scatter correction = yes, default)

$q = K_\text{pri}$ is scored directly (latch $=$ 0 photons).
$K_\text{pri}$ is read straight from the scoring array, so

$$R = \frac{K}{q}$$

Applying the delta method to $R = K/q$:

$$\left(\frac{\sigma_R}{R}\right)^2 = \delta_K^2 + \delta_q^2 - 2 \cdot r_c$$

where $\delta_X = \sigma_X / \bar X$ is the relative standard deviation
and $r_c = cov(K,q)/(\bar K \bar q)$ is the
relative covariance (not the Pearson coefficient; unbounded).

Scatter-dominant mode (scatter correction = no)

$q = K_\text{scat}$ is scored (latch $\neq$ 0 photons).
$K_\text{pri} = K - K_\text{scat}$, so

$$R = \frac{K}{K - q}$$

Applying the delta method to $R = K/(K-q)$:

$$\left(\frac{\sigma_R}{R}\right)^2 = (R-1)^2 \left(\delta_K^2 + \delta_q^2 - 2 \cdot r_c\right)$$

where $r_c = cov(K, K_\text{scat})/(\bar K \bar K_\text{scat})$.

The $(R-1)^2$ prefactor makes this mode dramatically less efficient at
depth: at 30 mfp in iron $R \approx 86$, so even near-perfect
correlation between $K$ and $K_\text{scat}$ is insufficient to overcome
the $\approx 85\times$ amplification of any residual decorrelation.
Benchmark runs show scatter-dominant mode is ${\sim}5.7\times$ less
efficient per history than scatter-correction mode at 30 mfp.

Note on $r_c$

$r_c$ is the relative covariance $cov(K,q)/(\bar K \bar q)$,
estimated from the cross-term array as

$$r_c = \frac{\overline{Kq} - \bar K \bar q}{\bar K \bar q} = \frac{\sum_i K_i q_i / N - \bar K \bar q}{\bar K \bar q}$$

It is not the Pearson correlation coefficient and is not bounded to
$[-1,1]$. Do not clamp it.

Input

:start scoring options:

    score primaries    = yes
    scatter correction = no    # default yes; set no for Kscat mode
    ...
:stop scoring options:

Implementation notes

• scg_kp is allocated only when score primaries = yes; fully
  serialized in outputData / readData / addState / resetCounter
  for correct parallel job combination.
• kerma_p stores either $K_\text{pri}$ or $K_\text{scat}$ depending
  on is_scatter_correction; the latch condition at scoring time is
  !latch (correction mode) or latch != 0 (scatter mode).