head()/tail() for emmGrid: follow base-R negative-n semantics

[SchmidtPaul]

In base R, head(x, -k) drops the last k elements and tail(x, -k) drops
the first k. The emmGrid methods have these swapped for a negative n:
head.emmGrid(x, -k) drops the first k rows and tail.emmGrid(x, -k) drops the
last k. Positive n is fine.

warp.rg <- ref_grid(lm(breaks ~ wool * tension, data = warpbreaks))  # 6 rows: L L M M H H (tension)

head(1:6, -2)   # base-R reference: 1 2 3 4  (drop last 2)
tail(1:6, -2)   # base-R reference: 3 4 5 6  (drop first 2)

# --- before:
as.character(head(warp.rg, -2)@grid$tension)   # "M" "M" "H" "H"   (dropped the FIRST 2)
as.character(tail(warp.rg, -2)@grid$tension)   # "L" "L" "M" "M"   (dropped the LAST 2)

# --- after:
as.character(head(warp.rg, -2)@grid$tension)   # "L" "L" "M" "M"   (drops the last 2)
as.character(tail(warp.rg, -2)@grid$tension)   # "M" "M" "H" "H"   (drops the first 2)

Fix: rewrote both methods in R/rbind.R to compute the kept-row count exactly
the way utils:::head.default() / tail.default() do — max(N + n, 0L) for a
negative n — using positive indexing only. n = 0 and abs(n) > nrow return
an empty grid, matching base behavior; positive n is unchanged. Present since
the head()/tail() methods were added (commit 3a63e5d, 2022-08-30; first
shipped in 1.8.1); the help page has no negative-n example, so no documented
behavior changes. The only internal callers use positive n, so package
internals are unaffected. Added a regression test in
tests/testthat/test-head-tail.R.

Found via an AI-assisted bug hunt (see #580).