feat: opt-in field filtering for scrape_batch and search_engine_batch

[SuarezPM]

Splitting the field-filtering part out of #140 into its own PR.

Adds an optional fields param to both batch tools so you only get back the
fields you actually need — saves a lot of tokens in agent pipelines.
filter_fields skips prototype-polluting keys and has unit tests.

• Backward compatible: if fields isn't passed, the output is exactly the
  same as before.
• scrape_batch also isolates errors per URL now (a failed URL comes back as
  {url, error} instead of breaking the settlement) and keeps
  responseType: 'text'.

Leaving the dedup out of this one — it needs a block-level rewrite to actually
collapse shared boilerplate properly, so that'll come as its own PR.

[egoriklok]

Public no-secret MCP Buyer-Agent Readiness Snapshot for brightdata-mcp:

• Public signal: Public MCP candidate brightdata-mcp: A powerful Model Context Protocol (MCP) server that provides an all-in-one solution for public web access.. Matched R1 terms: agent, automation, browser, github, mcp, server, tool.
• R1 fit: agent, automation, browser, github, mcp, server, tool.
• Readiness status: public evidence review needed before an autonomous buyer-agent should rely on this surface.
• Blind spot 1: explicit auth scopes and delegated-permission boundary.
• Blind spot 2: spend/API cost cap plus approval semantics before paid actions.
• Blind spot 3: receipt, audit-log, revocation, or dispute evidence for safe buyer-agent use.
• Single next question: For brightdata-mcp, is there already a documented policy for agent spend/auth limits, receipt evidence, and revocation before a buyer-agent can invoke it?

No secrets, invoice, payment link, delivery link, private endpoint, paid call, or wallet signature; this is only a free public snapshot.