r.sim: disable broken nblock usage

[petrasovaa]

TL;DR

Disabled the broken nblock >1 auto-split in r.sim.water/r.sim.sediment by forcing nblock=1. Kept the variables so that nblocks functionality can be enabled in follow up PRs.

Fixed the err formula to the original fortran's Monte Carlo stddev (fixes #4799). As a result error map is now 0 and will become more useful once nblocks will be enabled to be > 1.

Background

nblock originated in the Fortran ancestor as a memory workaround for running more walkers than the static buffer held; the C port inherited the structure but eventually mangled the per-block walker
accounting (rwalk was never actually divided), silently producing biased depth/discharge for users beyond the MAXW cap.

Summary of changes

• Disable the broken auto-split path in r.sim.water / r.sim.sediment: force nblock = 1, drop the MAXW static cap, and remove the obsolete "ERROR: nwalk > maxw" doc reference.
• Repair the err output to the Fortran's Monte Carlo standard-deviation formula (sqrt(|E[(gama·conn)²]/nblock − gama²|)); previously the C port had a wrong exponent and a missing
  post-loop step that produced a quantity with no clean meaning (fixes [Bug] r.sim.water: error output is not error but sum of depths? #4799). With nblock = 1 today it's zero everywhere; it becomes meaningful when nblock > 1 is reintroduced.
• Introduce conn = nblock/iblock to gama in output_data for depth, discharge, conc, and flux outputs, matching the Fortran's intermediate-snapshot extrapolation. With nblock = 1 conn = 1.0
  and outputs are unchanged.
• Preserve nblock, conn, and the iblock loop as scaffolding for sequential-replica reintroduction (Fortran-faithful, with working variance estimator) followed by thread-parallel replicas.

Behavior changes

• nblock = 1 users (the only path that actually worked): depth, discharge, conc, flux, erdep unchanged. err output changes from pow(gama_final, 3/5) (which had no clean physical meaning)
  to all-zeros (correct stddev for one sample).
• Users beyond the old MAXW cap: previously got silently biased outputs from the broken auto-split

Not in this PR

• Reintroducing nblock > 1 (sequential-replica execution) — to follow.
• Thread-parallel replica execution and the OpenMP race in the per-walker loop — separate follow-up.

[metzm]

The changes look good to me, thanks!

Looking forward to a proposal enabling nblocks > 1.