perf(build): optimize deps but not workspace crates in dev/test

[mohammadfawaz]

What

Speeds up local build iteration by splitting optimization between Leo's own crates and its dependencies:

• dev/test profiles: workspace crates drop from opt-level = 2 → 1 (fast incremental rebuilds), while all dependencies build at opt-level = 3 via [profile.<p>.package."*"]. snarkVM and friends are compiled once and cached, so runtime — including snarkVM proving in the test suite — stays fast even though our own crates are barely optimized.
• test profile now emits debug = "line-tables-only" instead of full debuginfo — keeps file/line in backtraces, much cheaper to generate and link.
• ci profile pins package."*" back to opt-level = 1, so the new dependency optimization does not leak in through inherits = "test" and slow CI's near-clean builds.

Why

The runtime hot path is inside snarkVM, not Leo's passes, so optimizing the workspace at -O2 on every edit cost compile time without buying meaningful runtime. This is the standard "optimize deps, not your code" dev-profile split.

Verification

cargo build -v confirms the resolved flags:

• dev: leo_span → opt-level=1, fxhash/indexmap → opt-level=3
• ci: every crate → opt-level=1 (CI compile time unchanged)

Deliberately not in this PR (need benchmarking / sign-off)

• Gating snarkVM's rocks feature off the default/test builds — would remove the RocksDB C++ + bindgen compile (and the CI libclang install) from builds that don't need persistent storage. Biggest cold-build win, but a behavior change.
• Shrinking the full-snarkvm reach in parser/ast (they pull the umbrella crate but mostly need the lighter snarkvm-console subset they already depend on) — improves build-graph parallelism, but needs an import refactor + verification.
• Faster linker (lld/mold) via .cargo/config.toml — left as a per-developer opt-in rather than forced repo-wide, since not every machine/CI image ships it.
• The existing .cargo/config.toml has a malformed cfg(... not(feature = "noconfig))) (unterminated string) that silently disables its target-cpu=native rule — flagging separately since fixing it changes runtime codegen, not build speed.

🤖 Generated with Claude Code