Fix Windows ACP install: spawn powershell.exe directly instead of via bash

[h4root]

Fixes #2401, fixes #2407.

Problem

On Windows, the CLI install commands for the Claude and Codex ACP runtimes are PowerShell one-liners (see managed_agents/discovery.rs):

powershell.exe -NoProfile -ExecutionPolicy Bypass -Command "irm https://claude.ai/install.ps1 | iex"

run_install_command currently wraps every install command in bash -l -c "..." via install_shell_command. This has two consequences on Windows:

1. Broken nested quoting — the inner double quotes around the -Command argument collide with the outer bash -c "..." quoting, so the PowerShell one-liner is mangled.
2. Git Bash required to install the CLI — resolve_install_shell returns GIT_BASH_INSTALL_HINT (no suitable shell found) when Git Bash is absent, so the user cannot install the Claude/Codex CLI at all until they first install Git Bash.

The bash -l login-shell startup also reads the user's full profile (nvm/conda/etc.), which is what produced the "took 45 minutes" report in #2401 before the 5-minute timeout killed it.

Fix

Add build_install_command, which on Windows detects powershell.exe ... commands and spawns powershell.exe directly (with CREATE_NO_WINDOW), bypassing Git Bash for that step. Everything else — the npm adapter installs, all Unix — continues through the existing install_shell_command path unchanged.

• PowerShell CLI installs no longer need Git Bash and no longer pay the bash -l startup cost.
• #[cfg(windows)] gates the whole addition, so macOS/Linux behavior is untouched.
• Unit tests cover detection (PowerShell yes; npm / curl-pipe-bash no; case-insensitive).

[h4root]

Closing — this is already fixed upstream in #2680 (is_powershell_command / install_powershell_command). This branch was based on a stale main and reimplements the same fix, less completely. Apologies for the noise.