docs(glimmer-attribute-behavior): rewrite from verified empirical data

Replaces the prior tables (which mixed verified data with extrapolations
marked "(assumed)") with strictly-verified per-attribute tables. Every cell
populated from rendering and IDL inspection in ember-source 6.12.

Structure:
- One "Reference table" section, five per-attribute sub-tables
  (muted, disabled, aria-hidden, tabindex, autocomplete)
- One "To reproduce the reference table" section with the exact template
  and JS console snippet, inline (no separate fixture file)
- Cross-attribute observations summarizing the rules the data reveals

Findings the new tables make explicit:
- Bare-mustache falsy set is {{false}}/{{null}}/{{undefined}}/{{0}} for
  boolean-coerced attrs (boolean HTML, ARIA, numeric); {{""}} is kept as
  attr="".
- Bare-mustache string literals never coerce — attr={{"false"}} renders
  as attr="false".
- Concat-mustache for boolean HTML attrs sets the IDL property to true
  regardless of the literal value inside (verified for both reflecting
  and non-reflecting attrs).
- Concat-mustache for ARIA/string attrs renders the stringified value
  literally — no boolean coercion. aria-hidden="{{false}}" is visible.
- Plain string attrs (autocomplete) skip Glimmer's boolean coercion
  entirely; autocomplete={{false}} renders as autocomplete="false".

The video.muted snapshot reads IDL muted=false for static attribute forms
(m1-m4, m7-m8, m11) because the test runs before media load — the doc
explains how defaultMuted reflects to muted at load time, so the rule's
"At play time" column is the lint-truth column rule authors should use.

Author: johanrd