Fix operator precedence bug in PaidIncurredChain and improve code quality

[fabioconcina]

Summary

• Fix operator precedence bug in PaidIncurredChain() that inflated reserves by ~2.9%
• Improve code quality: rename variables shadowing base R, pre-allocate vectors, use idiomatic R, remove dead code
• Add regression test

Bug details

R's : operator binds tighter than -, so tau2.hat[i:J-1] evaluates as tau2.hat[(i:J)-1] instead of the intended tau2.hat[i:(J-1)]:

# i=1: indices (1:J)-1 = 0:(J-1), index 0 silently dropped → correct by accident
# i=2: indices (2:J)-1 = 1:(J-1) → includes tau2.hat[1], should start at [2]

This caused v2[1] == v2[2], violating the strictly decreasing sequence required by Proposition 2.2 in Merz & Wüthrich (2010). Fixed with a vectorized cumsum:

v2 <- sum(sigma2.hat) + c(rev(cumsum(rev(tau2.hat))), 0)

Impact on USAA example data

Metric	Before	After	Delta
Total reserve	1,643,953	1,596,954	-2.9%
s.e.	113,940	110,977	-2.6%

Code quality improvements

• Renamed c → c_par, diag/diag2 → diag_I/diag_P (shadowed base R functions)
• Pre-allocated vectors instead of growing with c()
• Replaced variance loops with apply(), cumulative sum loop with cumsum()
• Adopted checkTriangle() and getLatestCumulative() (consistent with rest of package)
• Vectorized fP and fI triangle construction
• Simplified output to single list() call
• Removed 257 lines of commented-out Bayesian model code (preserved in git history)

Test plan

• New test.PaidIncurredChain passes (6 checks)
• Full test suite: 70 tests, 0 failures (3 pre-existing errors unrelated)