feat(agents): make reqwest/skills feature-driven and isolate ollama provider path

[M09Ic]

Background / Motivation

We need a minimal moltis agent footprint for edge/embedded environments. Today reqwest pulls a heavy HTTP/TLS stack into all agent builds.

The goal of this PR is to make that stack feature-driven while preserving default behavior, so minimal consumers can build an agent path without HTTP dependencies and later plug in a non-HTTP transport/provider.

In local minimal-path validation (cerebellum release), binary size dropped from 8,167,936 bytes to 5,128,192 bytes (-37.22%).

What This PR Changes

• Make reqwest optional in moltis-agents.
• Gate HTTP-dependent provider paths behind provider features.
• Keep default behavior compatible by enabling the existing HTTP provider set in default features.
• Keep skills/install forwarded via skills-install feature (enabled by default for compatibility).
• Extract Ollama live-model discovery into providers/ollama.rs and gate it via provider-ollama.

Design Direction (Important)

This PR is a prerequisite for a transport-agnostic architecture:

• moltis-agents can run with HTTP provider features disabled.
• External systems can provide model access through non-HTTP channels (e.g. socket/packet/RPC bridges).
• The non-HTTP provider implementation itself is not included in this PR.

Compatibility

• Default builds keep current behavior.
• Minimal builds can now selectively disable HTTP-related features to reduce dependency and binary footprint.

[M09Ic]

Edge Deployment Vision for moltis-agents

Objective

Enable a minimal, transport-agnostic agent runtime that can execute on constrained edge devices while reusing the existing moltis-agents core.

Proposed Architecture

• cerebellum is a thin adapter crate around moltis-agents (no agent reimplementation).
• Request/response payloads remain OpenAI-compatible at the schema level.
• Transport is a persistent socket channel (not HTTP), so minimal builds can avoid pulling in the full HTTP client stack.

Bridge Server Model

A lightweight bridge service (Python/Go/Rust/etc.) can:

1. terminate the socket protocol from cerebellum,
2. translate/forward calls to upstream providers (OpenAI, Anthropic, etc.),
3. stream tokens/events back over the same long-lived channel.

Expected Outcome

• Target ~1 MB agent binary in a minimal profile.
• Preserve long-lived bidirectional communication for dialogue, tool calls, and task orchestration.

Scope Note

This PR focuses on dependency shaping and feature-gating needed for this direction.
The custom non-HTTP provider implementation is intentionally out of scope here.

Maintainer note: from an engineering perspective, moltis has one of the cleanest architectures I have seen in this category, with strong long-term extensibility potential.