Feature: Support .env file loading and per-agent environment variable scoping

## Summary

Copilot CLI should support automatic `.env` file loading from the project root, and allow agents to declare environment variables scoped to their configuration — so that different agents can have isolated secrets without polluting the global shell environment.

## Problem

When building custom agents with Copilot CLI that use MCP servers requiring credentials (e.g., the Slack MCP server needing `SLACK_USER_OAUTH_TOKEN`), all credentials must be exported in `~/.bashrc` or equivalent shell profile because:

1. Copilot CLI does not load `.env` files from the project directory
2. Agent `env:` blocks in MCP config use `${VAR}` syntax, which resolves from the shell environment only
3. There is no way to scope environment variables per-agent

### No project-scoped configuration
Secrets for one project leak into every shell session. A `.env` file at the repo root (gitignored) is the standard pattern for project-scoped secrets in every major framework (Node.js, Python, Ruby, Go). Copilot CLI is the only tool in my workflow that forces secrets into the global shell profile.

### No secret isolation between agents
If you have multiple agents, each needing different MCP servers, every agent currently sees every exported variable in the shell. Per-agent env scoping would follow the principle of least privilege — an agent that only needs a Slack token should not have access to other API keys.

### Onboarding friction
When sharing a multi-agent repo, contributors need to manually add exports to their shell profile. A `.env.example` + `.env` pattern is universally understood and self-documenting.

## Proposed Solution

### Part 1: `.env` file loading (minimum viable)
- On startup, if a `.env` file exists in the project root (or git root), load it into the environment before resolving `${VAR}` references in agent/MCP configs
- Respect `.env` files at both user level (`~/.copilot/.env`) and project level (`.env` in git root), with project-level taking precedence

### Part 2: Per-agent environment scoping (stretch goal)
Allow agents to declare their own env vars or reference a specific `.env` file:

```yaml
---
name: slack-agent
env:
SLACK_USER_OAUTH_TOKEN: ${SLACK_USER_OAUTH_TOKEN}
env_file: .env.slack
---
```

This would restrict which variables are available to that agent's MCP server subprocesses.

## Current Workaround

Export all variables in `~/.bashrc` and maintain a `.env.example` in the repo for documentation only:

```bash
# ~/.bashrc
export SLACK_USER_OAUTH_TOKEN=xoxp-...
```

This works but violates standard project-scoping conventions.

## Related Issues

- #1232 — `${}` env var expansion in MCP config (implemented, but only reads from shell env)
- #1354 — Per-agent model selection (similar theme of per-agent configuration)

La payload mentionne l’initialisation de l’environnement de démarrage et la résolution de `${VAR}` dans la configuration de l’agent et de MCP, mais il n’y a ni fichiers d’implémentation ni tests. Commencez par suivre ces points d’entrée et la manière dont les configurations du projet et de l’utilisateur sont découvertes. Le travail doit inclure le comportement de priorité `.env` demandé ainsi qu’une conception ou une implémentation claire pour le périmètre par agent, avec une couverture des deux parties.

Rédigé par le modèle d'indexation à partir du texte de l'issue.